A1210
Title: Hypothesis testing for mediation effects in a generalized regression model
Authors: Jung Hyub Lee - The University of Texas at Austin (United States) [presenting]
Abstract: A unifying framework is considered for testing causal mediation effects in nonlinear models. A generalized linear-index model is introduced and extended to incorporate endogenous treatments and endogenous mediators. This model does not impose parametric assumptions on the error terms. A kernel-weighted Kendall's tau is leveraged to test the significance of the indirect effect of endogenous treatments on the outcome variable of interest mediated by endogenous mediators. The proposed semiparametric model allows for treatments and mediators to be discrete, continuous, or neither of these two (e.g., censored or truncated). Two distinct kernel-weighted Kendall's tau statistics will be constructed that capture the effect of the treatment on the mediator, and the mediator on the outcome variable of interest, respectively. However, it turns out that typical joint hypothesis tests using these statistics demonstrate the severely low size of a test and low test power. A similar problem also has been reported when using standard linear causal mediation models. To tackle the problem, a test method is leveraged that is a 'nearly similar powerful' that gives a size of the test sufficiently close to the desired level. For empirical illustration, we consider the British Household Panel Survey data to assess the effect of education level on social functioning mediated by annual individual income.
