Monday, Aug 4: 10:30 AM - 12:20 PM
4054
Contributed Papers
Music City Center
Room: CC-209C
In this session, presenters will demonstrate latest advancements in modeling in causal inference as well as handling unmeasured confounding.
Main Sponsor
Biometrics Section
Presentations
Mediation analysis plays a vital role in causal pathway studies which aim to determine whether the exposure variable affects the outcome through a third variable. Statistical tests for mediation analysis examine the presence of mediation effects, such as how the environment may impact one's behavior by influencing specific areas of the brain. However, traditional testing methods have been shown to be underpowered near the origin as they fail to specify the accurate null hypothesis, which is a combination of three different types. To fix this problem, we propose an adaptive test method by introducing an adaptive operator toward the original test statistic. Through both theoretical and simulation results, we will show that our method is able to control type I error properly while enhancing power compared to traditional test methods. A real data experiment based on the ABCD dataset will also be performed to demonstrate the improvement of our testing method.
Keywords
Causal inference
Mediation analysis
Sobel test
Adaptive test
Clustered time-to-event data are increasingly common in health research, particularly in observational studies such as with electronic health records or administrative data. For example, time-to-event data are often collected from individuals within shared environments such as hospitals, which may result in interdependence between observations. To quantify treatment effects from observational data we typically rely on causal inference approaches such as propensity score weighting to control for confounding. However, despite extensive literature on causal approaches for time-to-event data and clustered data individually, methods for incorporating both remain understudied. Motivated by this methodological gap, we extend existing causal survival estimators to accommodate clustered data and investigate their performance in a comprehensive simulation study with emphasis on varying clustering perspectives and robustness to misspecified propensity score and censoring score models. We lastly provide an accessible R tutorial to demonstrate how to implement our estimators with a real-world clinical example as to provide practical guidance for applied researchers interested in clustered causal survival analysis.
Keywords
Biostatistics
Causal Inference
Survival Analysis
Longitudinal/Correlated Data
Random Effects and Mixed Models
Electronic Health Records
Causal path analysis, an extension of causal mediation analysis, seeks to decompose a treatment or exposure effect into multiple path-specific effects. One particular methodology, generalized causal mediation and path analysis, allows for discrete or continuous mediators (and final outcome) and handles path models with multiple causally-ordered sets of mediator. As this method, which uses an extended mediation-formula for estimation, makes strong sequential ignorability assumptions, it is important to accompany analyses with relevant sensitivity analyses. The present work presents and explores a flexible approach to sensitivity analysis, based on a regression relationship between potential outcomes for pairs of model variables with possible unmeasured confounding. We provide simulation study results showing good properties for path-specific effect estimates under the extended (sensitivity analysis) model, and expound on the interpretation of the sensitivity parameter. Using our recently developed GMediation R package, we illustrate the new methods 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
Co-Author(s)
Carly Rose, Case Western Reserve University
Fang Wang, Case Western Reserve University
Yifei Xu, Case Western Reserve University
Ming Wang, Case Western Reserve University
Jang Ik Cho, Meta Platforms Inc., Reality Lab Health Technology
First Author
Jeffrey Albert, Case Western Reserve University
Presenting Author
Jeffrey Albert, Case Western Reserve University
Unmeasured confounding challenges causal inference in intensive longitudinal studies, potentially biasing treatment effect estimates. The proximal causal learning framework offers a promising approach to nonparametric identification using proxies or negative control variables in the presence of hidden confounding bias. While prior literature considers the joint effect of time-varying treatments, our work extends the framework to a time series setting to estimate the contemporaneous effect of time-varying treatments. We demonstrate that under traditional proximal causal inference assumptions, we can recover unbiased effect estimation in the presence of unmeasured confounding by leveraging the intensive longitudinal nature of time series data. Specifically, we use past and future observations as natural proxies for unmeasured confounders and revise the bridge function for valid estimation. Simulation studies illustrate our method's potential for studying environmental science and wearable device data. Our work contributes to the growing literature on proximal causal inference and provides a powerful tool for analyzing longitudinal data, including in mobile health research.
Keywords
unmeasured confounding
proximal causal inference
time-series
mhealth
Data-driven personalized decision-making has become increasingly important in recent statistical research. Existing methods often rely on the assumption of no unmeasured confounding to establish valid causal inferences before proceeding with decision-making for identifying the optimal individualized treatment rule (ITR). However, this assumption cannot be guaranteed in practice, especially in observational studies. While additional data sources such as instrumental variables or proxies have been commonly utilized to address unmeasured confounding, such information is not always available. In this work, we develop a novel Bayesian approach for robustly learning the optimal ITR under unmeasured confounding. We propose a Bayesian joint model for continuous outcome and treatment, accounting for observed covariates and unmeasured confounders. We prove that the proposed joint model achieves unique causal identification under certain mild distributional assumptions, without requiring additional data sources. Through simulation studies and an application to a large-scale kidney transplantation database, we demonstrate the identifiability, utility, and robustness of the proposed method.
Keywords
Causal identification
Conservative policy optimization
Individualized treatment rules
Precision medicine
Unmeasured confounding
Co-Author(s)
Yang Ni, Texas A&M University
Yanxun Xu, Johns Hopkins University
First Author
Wei Jin, Johns Hopkins University
Presenting Author
Wei Jin, Johns Hopkins University
When multiple outcomes are of interest in an observational study, each outcome often exhibits different sensitivities to unmeasured confounding, where some outcomes are more sensitive to biases from an unmeasured confounder while others are less sensitive. Also, if the outcomes share a common low-dimensional structure the individual biases from unmeasured confounding also have (up to rotation) the same low-dimensional structure. Leveraging both features, we propose a novel procedure, LaunchODR, which conducts a sensitivity-aware dimension reduction for testing causal effects in observational studies with multivariate outcomes. Specifically, LaunchODR modifies standard dimension reduction methods to identify the shared low-dimensional structure and conducts the "least sensitive test" to assess average treatment effects across outcomes. Under some assumptions, we show LaunchODR asymptotically identifies the correct low-dimensional embedding, and the resulting test has Type I error control and consistency. We demonstrate LaunchODR on a population-scale epigenomic dataset to study the causal effect of Amyloid-β on multiple DNA methylation regions.
Keywords
Sensitivity analysis
Unmeasured confounding
Dimension reduction
epigenomics
DNA methylation
Alzheimer's disease
Unmeasured confounding presents a significant challenge in causal inference from observational studies. Classical approaches often rely on collecting proxy variables, such as instrumental variables. However, in applications where the effects of multiple treatments are of simultaneous interest, finding a sufficient number of proxy variables for consistent estimation of treatment effects can be challenging. Various methods in the literature exploit the structure of multiple treatments to address unmeasured confounding. In this paper, we introduce a novel approach to causal inference with multiple treatments, assuming sparsity in the causal effects. Our procedure autonomously selects treatments with non-zero causal effects, thereby providing a sparse causal estimation. Comprehensive evaluations using both simulated and Genome-Wide Association Study (GWAS) datasets demonstrate the effectiveness and robustness of our method compared to alternative approaches.
Keywords
Causal inference
Identification
Sparsity assumption
Unmeasured confounding.