A0168
Title: Nonparametric mediation analysis: Beyond the mean
Authors: Yen-Tsung Huang - Academia Sinica (Taiwan) [presenting]
Abstract: Mediation analyses estimate the effect of exposure on an outcome mediated by a mediator. Many methods have been developed to conduct mediation analyses, and most of them focus only on the mean outcome. The biological effect of a cancer mutation may have affected the variation of a downstream outcome, such as gene expression and patient survival, without even changing the mean outcome. To this end, the effect on the cumulative distribution function (cdf) of an outcome is characterized. Consequently, one can easily summarize the impact of an outcome on any specific moment, with the effect on the mean outcome as a special case. A nonparametric estimator is proposed based on kernel estimators for the cdf of the mediator given the exposure and that of the outcome given the exposure and mediator. The uniform consistency and weak convergence of the proposed estimators are established. Extensive simulation studies were conducted to evaluate the performance of finite samples. These methods are applied to two studies: one investigates the influence of childhood socioeconomic adversity on adult adiposity via DNA methylation of the FASN (fatty acid synthase) gene, and another investigates how IDH1 (isocitrate dehydrogenase 1) mutations in glioma patients affect EGFR (epidermal growth factor receptor) expression by altering its DNA methylation.