Correlation is not causation. As such, there are decision-making contexts (like Newcomb's Problem) where it is entirely reasonable for an agent to believe that the world is likely to be better when she x's, but also believe that x-ing would cause the world to be worse. Should agents x in such contexts? In this paper, I use the interventionist approach to causation to help answer this question. In particular, I argue that whether an agent should x depends on her credence that her decision constitutes an intervention. I also propose and defend a decision rule that takes stock of the exact way in which what an agent should do depends on this credence.