In 2000, Kent Beck popularized the notion of pair programming, and argued that this practice could improve software quality and developer learning, and would not adversely affect duration, as compared to solo programming. We conduct a meta-analysis of 15 years of empirical tests of these relationships, and find that: pairs generally (a) produce higher quality software, (b) learn more, and (c) program faster, supporting Beck's arguments. We conclude the paper with what we know and what we do not know about pair programming and directions for future research on the topic.