This paper presents the first analysis of the life course outcomes through late midlife (around age 55) for the participants of the iconic Perry Preschool Project, an experimental high-quality preschool program for disadvantaged African-American children in the 1960s. We discuss the design of the experiment, compromises in and adjustments to the randomization protocol, and the extent of knowledge about departures from the initial random assignment. We account for these factors in developing conservative small-sample hypothesis tests that use approximate worst-case (least favorable) randomization null distributions. We examine how our new methods compare with standard inferential methods, which ignore essential features of the experimental setup. Widely used procedures produce misleading inferences about treatment effects. Our design-specific inferential approach can be applied to analyze a variety of compromised social and economic experiments, including those using re-randomization designs. Despite the conservative nature of our statistical tests, we find long-term treatment effects on crime, employment, health, cognitive and non-cognitive skills, and other outcomes of the Perry participants. Treatment effects are especially strong for males. Improvements in childhood home environments and parental attachment appear to be an important source of the long-term benefits of the program. (author abstract)