Causal Mediation: Assessing Meaningful Path-specific Effects Rather than Decomposing the Total Effect

Nonparametric causal mediation methods (Pearl, 2001; Robins and Greenland,1992) are commonly described as decomposing the total effect into thenatural direct and indirect effects. Here, we argue that mediation should not beunderstood in terms of decomposition, as there is no unique way to partition atotal effect into additive contributions of direct and indirect paths outside thelinear context (Baron and Kenny, 1986). Instead, the use of mediation methodsshould be guided by the specific research questions one seeks to address. Wediscuss how clearly formulated research questions about the mediation processnaturally connect to path-specific direct and indirect effects. Once the priorityof research questions is established, the practical irrelevance of statisticaleffect decomposition directly follows. Finally, we contrast natural effects withalternative approaches and explore their implications for recent applications ofmediation to questions about discrimination and fairness.