Grant Number:	1R01CA311340-01 Interpret this number
Primary Investigator:	Derkach, Andriy
Organization:	Sloan-Kettering Inst Can Research
Project Title:	Mediation Analysis of Epidemiologic Data From Multiple Incomplete Data Sources
Fiscal Year:	2026

ABSTRACT This project concerns how to perform mediation analysis when we have three ‘incomplete’ data sets, each containing measurements for only two of the relevant variables. In recent years, mediation analysis has become a commonly used tool in epidemiologic studies to decipher the role of a variable (or a set of variables) in explaining a well-established relationship between an exposure and an outcome. In parallel, many advanced methods for mediation analysis have been proposed to handle many different data types and configurations. However, to-date, all methods require the three relevant variables to be measured in a single, common data set; thus, they are not suitable for the mediation analysis with multiple data sources when there is no single resource that contains joint information on all key variables. Our proposal is motivated by our observation that there is a growing need for a statistical framework for conducting mediation analysis in this context. Inspired by past and ongoing collaborations, we propose to develop novel general statistical tools that can handle different combinations of data sets with various levels of information. In Aim 1, we will propose an efficient semiparametric framework for estimating regression parameters without specifying parametric distributions of the exposure and the mediator. This aim will build upon preliminary work based on the assumption that the multiple data sources are arising from identical underlying populations. Methodology in Aim 2 will relax the assumption that multiple sources are arising from identical underlying populations. We will propose new framework to account for heterogeneity in distributions of covariates, mediators and exposures between three sources. In Aim 3, we will present a novel methodology for high-dimensional mediation analysis by building on the developments in Aims 1 and 2. We will introduce novel methods that incorporate L1 penalties and latent variables to the analysis without requiring individual-level data. For the settings examined in all aims, we will compare the efficiencies of estimators obtained under various study designs to provide important guidance for planning of future epidemiological studies. The ultimate goal of the proposal is to create practical tools (see Aim 4) for modern mediation analysis that integrate data from multiple sources when there is no single resource that contains joint information on all key variables. Specifically, proposed software will have four sets of tools that provide power and sample size calculations, test sufficient conditions of identifiability of the regression and causal effects, estimate mediation effects and conduct sensitivity analysis.





None