An essential
goal of program evaluation and scientific research is the
investigation of causal mechanisms. Over the past several decades,
causal mediation analysis has been used in medical and social
sciences to decompose the treatment effect into the natural direct
and indirect effects. However, all of the existing mediation
analysis methods rely on parametric modeling assumptions in one way
or another, typically requiring researchers to specify multiple
regression models involving the treatment, mediator, outcome, and
pre-treatment confounders. To overcome this limitation, we propose a
novel nonparametric estimation method for causal mediation analysis
that eliminates the need for applied researchers to model multiple
conditional distributions. The proposed method balances a certain
set of empirical moments between the treatment and control groups by
weighting each observation; in particular, we establish that the
proposed estimator is globally semiparametric efficient. We also
show how to consistently estimate the asymptotic variance of the
proposed estimator without additional efforts. Finally, we extend
the proposed method to other relevant settings including the causal
mediation analysis with multiple mediators. (Last Revised January,
2016) |