Abstract: Heterogeneous effects are ubiquitous in the social sciences and are often an important component of---and means of assessing---theoretical arguments. A common strategy to assess this heterogeneity is to include a single multiplicative interaction term between the treatment and a hypothesized effect moderator in a regression model. In this paper, we show how inferences about interactions under this approach are highly sensitive to modeling choices about how the effect modifier interacts with other covariates, an issue almost never discussed in practice. We propose an alternative strategy that uses machine learning techniques to allow for more flexible estimation of interactions. We apply this approach to two applications: the effect of direct primary adoption on third-party voting, with heterogeneity by region, and the effects of remittances on political protest as moderated by level of democracy.