|
Abstract:
|
Causal mediation analysis plays an important role in program development to understand causal pathways through which an effect arises. To evaluate an indirect effect transmitted through a hypothesized mediator and a direct effect, causal mediation studies often take advantage of experimental designs. However, partial compliance has been common in experimental research. Given the post-treatment assignment nature of the treatment variable, the direct and indirect effects of the treatment assignment can not be identified without further assumptions beyond sequential ignorability. Under a deterministic principal stratification framework, Yamamoto (2014) proposed a local sequential ignorability assumption which identifies the direct and indirect effects of the treatment among compliers. However, the deterministic framework is incompatible with key concepts in causal mediation analysis. To reflect the stochastic reality of program participation and the intermediate experiences, we propose assumptions to decompose the effect of treatment among the “strength-of-IV weighted population” (Small et al, 2017) under a stochastic framework. We apply our method to National Job Corps Study.
|
