This paper proposes a general approach to causal mediation analysis that is not dependent on a specific statistical model. The approach defines causal mediation effects, establishes identification of these effects under an assumption of sequential ignorability, and develops general estimation procedures and sensitivity analyses that can be applied across different model types and data structures. The approach overcomes limitations of prior work that was based on linear structural equation models. The paper illustrates the approach using a job search intervention study and provides software implementing the proposed methods.