Statistical Inference for High Dimensional Mediation Effects

3431 

Contributed Abstract 

Poster 

Yuzhou Lin (1), Xihong Lin (2)

(1) Department of Statistics, Harvard University, Cambridge, MA, USA, (2) Harvard T.H. Chan School of Public Health, Cambridge, MA, USA

Xihong Lin  
Harvard T.H. Chan School of Public Health

Yuzhou Lin  
Department of Statistics, Harvard University

Yuzhou Lin  
N/A

Evaluating the effect of a treatment on an outcome via a mediator has received growing attention in clinical and genetic studies. Traditional mediation effect testing methods, including the Wald-type Sobel's test and the Joint Significance test, suffer from overconservative type-I-error and low power under a great quantity of composite null hypotheses. The recently developed divide-aggregate-composite-null test (DACT) properly controls the type-I-error with high power when any of its composite null case has proportion close to 1. But DACT's performance in other settings is unclear. We showed that under unfavorable settings, when no case has proportion close to 0 or when the effect size is large, DACT will fail to control the type-I-error, even with its default normal calibration under Efron's empirical null framework. We proposed a new calibration involving a three-component mixture model for DACT. We controlled the type-I-error while preserving high power compared with state-of-the-art testing methods under both favorable and unfavorable settings. A new procedure for estimating null proportions and a variation of DACT is proposed to boost its null estimation accuracy and power.

mediation effect|indirect effect|divide-aggregate composite-null test|mixture model|null proportion estimation|composite null hypothesis

Section on Statistics in Genomics and Genetics

Miscellaneous