The minimal model semantics is a natural interpretation of defaults yet it often yields a behavior that is too weak. This weakness has been traced to the inabiity of minimal models to reflect certain implicit preferences among defaults, in particular, preferences for defaults grounded on more ’specific' information and preferences arising in causal domains. Recently, ’specificity' preferences have been explained in terms of conditionals. Here we aim to explain causal preferences. We draw mainly from ideas known in Bayesian Networks to formulate and formalize two principles that explain the basic preferences that arise in causal default reasoning. We then define a semantics based on those principles and show how variations of the algorithms used for inheritance reasoning and temporal projection can be used to compute in the resulting formalism.