This paper uses population panel data from Sweden to investigate the causes and consequences of self-employment over the life cycle, and to evaluate how self-employment decisions can be influenced by policy. In the first part of the paper, I use machine learning methods to summarize the patterns of self-employment behavior observed in the data. I find that careers involving self-employment fit into a small number of economically distinct groups. Some self-employment spells are short, with minimal capital investment and rapid return to paid employment, while others persist and have substantial capital devoted to the business from the outset. Guided by these descriptive results, I develop and estimate a dynamic Roy model in which self-employment decisions depend on factors such as cognitive and non-cognitive skills, prior work experience, the cost of capital, and other labor market opportunities. The model integrates traditional models of dynamic career choice that feature human capital investment and models of business start-up that feature physical capital investment. I estimate the model and use it to evaluate policies designed to promote self-employment. Cognitive and non-cognitive skills, education, and prior work experience are important determinants of the types of businesses individuals start, how much capital they employ, and how long they remain in self-employment. Subsidies that incentivize self-employment are generally ineffective, both in terms of promoting long-lasting firms and in terms of improving the welfare and earnings of those induced to enter self-employment.