Nonparametric Partial Identification of Causal Net and Mechanism Average Treatment Effects

When analyzing the causal e§ect of a treatment on an outcome it is important to un- derstand the mechanisms or channels through which the treatment works. In this paper we study net and mechanism average treatment e§ects (NATE and MATE, respectively), which provide an intuitive decomposition of the total average treatment e§ect (ATE) that enables learning about how the treatment a§ects the outcome. We derive informative non- parametric bounds for these two e§ects allowing for heterogeneous e§ects and without re- quiring the use of an instrumental variable or having an outcome with bounded support. We employ assumptions requiring weak monotonicity of mean potential outcomes within or across subpopulations deÖned by the potential values of the mechanism variable under each treatment arm. We illustrate the identifying power of our bounds by analyzing what part of the ATE of a training program on weekly earnings and employment is due to the obtainment of a GED, high school, or vocational degree.