Causes do not invariably raise the probabilities of their effects, nor are they generally evidence for their effects. Learning about a cause can convey information about one of its effects either via a direct cause-to-effect inference, which will confirm the effect, or via a "backtracking" inference, which can disconfirm the effect by indicating that stronger inhibiting causes will occur. My aim is to find ways of separating the direct, or "front-door, evidence that causes provide for their effects in virtue of being causes from any backtracking evidence they might provide. In this way, I hope to salvage an important part of the idea that causes generally provide evidence for their effects. My argument relies heavily on the theory of causal Bayesian networks developed in J. Pearl's Causality and P. Spirtes, C. Glymour, and R. Scheines's Causation, Prediction and Search. In particular, I will show how to use Pearl's method of "adjustment for direct causes" to define functions that split the evidence that a cause provides for its effect into a "direct" and "backtracking" part (at least in the context of Markovian causal graphs). I will show that my use of this method does not commit me to any problematic "interventionist" metaphysics of causation of the sort that Pearl recommends. Indeed, I shall argue that a proper understanding of the epistemology of causation does not require a commitment to any specific metaphysics of the notion. If time allows I will discuss some of the limitations of the "causal net" approach in the context of my project, and will say something about how these limitations might be overcome. A complete copy of the talk is on the web at http://www-personal.umich.edu/~jjoyce.