Causal Mediation Analysis With A 3-Dimensional Image Mediator

In the study of psychology,neuroscience,and statistics,the research component of causal mediation analysis is usually to determine whether there are causal indirect effects between“a treatment” and “the potential outcome”,that is,does the treatment leads to the outcome through the mediator? In these researches,scholars often use structural equation models to estimate the parameters of interest in the causal mediation analysis.In the past literature,the data types of mediators that can be included in these models,including constants,scalars,and curves.However,with the widespread application of the magnetic resonance images in the fields of psychology and neuroscience,the work of introducing image variables into such structural equation models is imminent.To this end,in this paper we propose a novel method called image causal mediation analysis method,which can be used to estimate the causal indirect effects when the mediator is a three-dimensional image.The new method is constructed in two main parts: the first part is to construct the linear image structural equation models,which have a three-dimensional image mediator,and to design reasonable parameter identifiability assumptions for it,and the second part is to establish an effective parameter estimation method.In order to verify the effectiveness of the proposed image causal mediation analysis method,a series of simulations are conducted in this paper.In addition,the proposed method is applied to observational study with a magnetic resonance brain image mediator,and some enlightening results are obtained.