Recent criticism of epidemiologic methods has focused on the limitations of 'black box' epidemiology, a pejorative label given to the simple identification of exposure-disease relationships. The assessment of mediation is an important tool for addressing this criticism. By using mediation analysis to open the black box, underlying mechanisms of the observed associations can be described and causal inference improved. An explicit theoretical motivation for such an analysis has been missing from the epidemiological literature. To provide this motivation, we integrate literature from epidemiology and other social sciences to describe the reasons that an investigator might want to assess mediation. We then describe the connections between these reasons and specific measures of indirect and direct effects that have been previously described.