This dissertation addresses the problem of when one can infer that a causal factor that has an effect in one population would have a similar effect in others. For example, if reducing class sizes increases educational outcomes in one neighborhood, would reducing them in another raise outcomes by a similar amount? This is known as the problem of extrapolation. This dissertation reviews the prior literature on extrapolation, explains how extrapolation relates to other forms of causal inference, and presents techniques for extrapolating more reliably. To model extrapolation, I rely on recently developed causal modeling techniques, which use graphs to represent causal relations among variables. I build on the work of Judea Pearl and Elias Bareinboim, who provide graphical methods for determining when it is possible to extrapolate causal quantities across populations given assumptions about how those populations differ. I argue that their account explicates one type of extrapolation, but that there are extrapolative inferences that go beyond their account. The central positive contribution of the dissertation is that it makes precise the sense in which knowing how a cause brings about its effect facilitates extrapolation. In cases where a cause influences its effect via multiple paths, newly developed “causal mediation techniques” enable one to precisely quantify the way that the cause influences its effect via each of the paths. These techniques aid extrapolation, since the causal quantities identified by these techniques are invariant across a range of ways that two populations may differ. Moreover, the conditions across which these quantities are invariant cannot be represented within Pearl and Bareinboim’s framework. In discussing the problem of extrapolation, I engage with several central philosophical issues. First, characterizing the problem requires one to elucidate the relationship between causal and statistical inference. Second, the causal mediation techniques I advocate shed light on recent debates about mechanistic explanation. Finally, the study of the conditions under which causal relationships generalize is essential for understanding the nature of causal relationships and their role in scientific theories.