Causal mediation analysis in mental health with mediator missingness

点击翻译按钮获取中文摘要

PROJECT SUMMARY Examining the mechanisms behind exposures and interventions is crucial in mental health research; doing so can help identify potential intervention targets and tailor interventions to optimize outcomes. Causal mediation analysis is an important tool for this job. An important but under-appreciated challenge in mediation analysis is missing data. Missing data are ubiquitous (even in high quality mental health studies) and are more of a challenge in mediation analyses than in standard analyses, due to the involvement of more variables. Missingness tends to be worse when the analysis involves multiple mediators and/or when data collection at intermediate time points (when mediators are measured) receives less attention than baseline and outcome data collection. Since most methodological work assumes complete data, it is not yet well understood how to appropriately handle missing data when conducting causal mediation analysis. As the first step in removing this barrier to effective use of causal mediation analysis in mental health researcher, this project will tackle the problem of mediator missingness, a common problem in practice and a particularly important one to tackle for the validity of mediation analyses. This project will develop two classes of methods for this purpose: (1) with the estimating equations approach, we will transform each of a collection of full-data methods to observed-data methods that are doubly robust – consistent if either a missingness model or a model for some relevant function of the mediator given observed data is correct; (2) to respond to the popularity of multiple imputation, based on each full-data method, we will develop targeted multiple imputation procedures tailored to the method and the effects being estimated, to minimize bias due to misspecification of the imputation model. The project seeks methods that can be explained intuitively, and will pay special attention to communicating and disseminating the methods developed to mental health researchers. For illustration, the methods will be applied to an analysis of effects of a treatment for anxiety disorder and an analysis of disparities in suicidality affecting sexual minority youth.