Sunday, Aug 4: 2:00 PM - 3:50 PM
5003
Contributed Papers
Oregon Convention Center
Room: CC-E143
Main Sponsor
Biometrics Section
Presentations
Mediation analysis plays a crucial role in causal inference as it can investigate the pathways through which the treatment influences the outcome. Most existing mediation analysis assume that the mediation effects are static and homogeneous within populations. However, mediation effects usually change over time and exhibit significant heterogeneity in many real-world applications. Additionally, the presence of unobserved confounding variables poses a significant challenge in inferring both causal effect and mediation effect. To address these issues, we propose an individualized dynamic mediation analysis method. Our approach can identify the significant mediators on the population level while capture the time-varying and heterogeneous mediation effects via latent factor modeling on coefficients of structural equation models. Another advantage of our method is that we can infer individualized mediation effects in the presence of unmeasured time-varying confounders. We provide estimation consistency for our proposed causal estimand and selection consistency for significant mediators. Extensive simulation studies and an application to DNA methylation study demonstrate the effectiveness and superiority of our method.
Keywords
Mediation analysis
Heterogeneous effects
Unmeasured confounding
Factor model
Structural equation model
Variable selection
Mediation analysis with one exposure (X) and one outcome (Y) is fairly well
developed. When there are multiple exposures, mediators, and outcomes, there
appear to be no standard concepts or formulas.
Also, if there are multiple exposures or mediators, the effect of one exposure
on an outcome may be partially mediated by the correlation of that exposure
with other exposures; the same is true of outcomes. This means that there may
be an effect of an exposure (X) or mediator (Z) on an outcome (Y) although a
linear model for Y may indicate that X or Z is not significant as a direct
covariate.
We have developed measures of direct and indirect effects of exposures on
outcomes in the situation where there are multiple exposures, mediators, and
outcomes, and all models are linear. The measures are derived from path
analysis as developed by Wright. We demonstrate with an example based on
simulated data.
Keywords
Mediation analysis
multiple covariates
path analysis
sparse data
Depression and Alzheimer's Disease (AD) are both prevalent diseases in older adults. Using the data sets from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study, we explore whether geriatric depression has a significant average treatment effect on AD and whether the effect is mediated by some important mediators. To estimate these causal effects consistently, we control for ultra-high dimensional potential confounders, including DNA methylation levels. We propose a new ball correlation-based screening method for confounder selection in mediation analysis. To achieve robustness against model misspecification, we utilize a robust mediation analysis framework. Simulation studies show that the proposed method has good finite-sample performance in terms of confounder and mediator selection, effect estimation, and inference. In the real data analysis, we find that geriatric depression has a significantly positive causal effect on AD. We also propose new prevention and treatment strategies for geriatric depression and AD through changing the selected confounders and mediators.
Keywords
causal inference
mediation analysis
Alzheimer’s disease
geriatric depression
ultra-high dimensional potential confounders
Co-Author(s)
Annie Qu, University of California At Irvine
Yubai Yuan, Pennsylvania State University
Qi Xu, University of California-Irvine
Fei Xue, Purdue University
Kecheng Wei, Fudan University
First Author
Yuexia Zhang, The University of Texas at San Antonio
Presenting Author
Yuexia Zhang, The University of Texas at San Antonio
We propose semiparametric Bayesian inference for causal mediation analysis in the context of cluster randomized trials (CRTs). Our approach allows for the estimation of direct and indirect effects at both the individual and cluster-levels. To model the joint distribution of cluster-level and individual-level confounders, we specify a two-stage Bayesian bootstrap (BB) with a "distance" metric between clusters, which avoids the need for restrictive parametric assumptions and allows us to borrow more information from "closer" clusters. By combining the observed data with causal assumptions, we are able to identify and estimate the natural direct and indirect effects at the individual-level and cluster-level separately. We assess the performance of our approach through simulation studies and use it to assess mediation in a cluster randomized trial in Kenya.
Causal mediation analysis seeks to decompose a treatment or exposure effect into direct and indirect effects. Initially developed for a single mediator, this methodology has been extended to allow for multiple contemporaneous or causally ordered mediators. Generalized causal mediation and path analysis (GCMPA) provides a further extension by accommodating multiple mediators at each of two causally-ordered 'stages' with mediators and the final outcome following generalized linear models. As these methods rely on the untestable assumption of sequential ignorability, a sensitivity analysis is widely considered to be an important accompaniment to causal mediation analysis. Unfortunately, methods for sensitivity analysis are currently available only for the single mediation case. We present a new sensitivity analysis approach applicable to the GCMPA setting, thus, evaluating the effect of departures from sequential ignorability on estimated path-specific effects. We discuss implementation of this methodology in the recently developed GMediation R package, and illustrate using data from a study of causal pathways between socioeconomic status and adolescent dental caries.
Keywords
causal inference
generalized linear models
path-specific effects
sequential ignorability
Our original work was motivated by the question of whether and to what extent well-established risk factors mediate the racial disparity observed for colorectal cancer incidence in the US. Typical mediation analysis examines the relationships between exposure, a mediator, and an outcome but requires access to a single complete dataset containing all three variables. However, because population-based studies include only a few participants from racial minority groups, these approaches have limited utility here. For this purpose, I developed novel methods to integrate several data sets with partial information for mediation analysis that accommodates complex survey and registry data and allows for multiple mediators. I then apply our method to data from US cancer registries, a US population-representative survey, and summary-level odds-ratio estimates of selected CRC risk factors from a case-control study. In this presentation, I will discuss several approaches to evaluate the robustness of results to violation of model assumptions.
Keywords
sensitivity analysis
data integration
summary level information,
survey sampling
registry data
First Author
Andriy Derkach, Memorial Sloan Kettering Cancer Center
Presenting Author
Andriy Derkach, Memorial Sloan Kettering Cancer Center
Causal mediation analysis is an important tool for investigating the causal effects of medications on disease-related risk factors, and on time-to-death (or disease progression) through these risk factors. However, such analyses are complicated by the longitudinal structure of the risk factors and the time-varying confounders. We develop a causal mediation approach, using (semi-parametric) Bayesian Additive Regression Tree (BART) models for the longitudinal and survival data. Our framework allows for time-varying exposures, confounders, and mediators, all of which can either be continuous or binary. We also quantify direct and indirect causal effects in the presence of a competing event. Motivated by data from the Atherosclerosis Risk in Communities (ARIC) cohort study, we use our methods to assess how medications, prescribed to target the cardiovascular disease (CVD) risk factors, affect the time-to-CVD death.
Keywords
causal mediation
longitudinal and survival data
semi-parametric
BART
cardiovascular