Tuesday, Aug 5: 2:00 PM - 3:50 PM
4117
Contributed Papers
Music City Center
Room: CC-207D
Main Sponsor
ENAR
Presentations
The direct and indirect effects of treatments or treatment strategies over time are often of interest in clinical research. For such questions, methodological challenges can include time-dependent confounding, informative timing of treatment decisions, and censoring. To address potential unidentifiability arising from post-treatment confounding and the right censoring, we define the mediator distribution using a random intervention (RI)-based conditional distribution. For modeling the observed data, we employ a generative Bayesian semiparametric survival model with a correlated gamma process hazard. The mediation effect is estimated through Monte Carlo g-computation using posterior draws from a blocked Metropolis-in-Gibbs sampler. We assess the performance of the method using simulated data, focusing on counterfactual effect predictions and mediation effects estimation under hypothetical treatment rules. We apply the method to a pediatric acute myeloid leukemia (AML) study to evaluate mediation effects of dynamic treatment regimens through clinical biomarkers such as organ failure.
Keywords
Causal mediation analysis
Bayesian survival analysis
dynamic treatment regimes
gamma process
proportional hazards model
Analyzing data from multiple sources offers valuable opportunities to improve the estimation efficiency of causal estimands. However, this analysis also poses many challenges due to population heterogeneity and data privacy constraints. While several advanced methods for causal inference in federated settings have been developed in recent years, many focus on difference-based averaged causal effects and are not directly applicable to multiplicative-scale estimands for a target population. More importantly, most methods are not designed to study effect modification. In this study, we introduce a novel targeted-federated learning framework to study the heterogeneity of treatment effects (HTEs) for a targeted population by proposing a projection-based estimand. This HTE framework integrates information from multiple data sources without sharing raw data, while accounting for covariate distribution shifts among sources. Our proposed approach is shown to be doubly robust and can conveniently handle continuous and categorical outcomes. Furthermore, we develop a communication-efficient bootstrap-based selection procedure to detect non-transportable data sources, thereby enhancing robust information aggregation without introducing bias. The superior performance of the proposed estimator over existing methods is demonstrated through extensive simulation studies, and the utility of our approach has been shown in a real-world data application using nationwide Medicare-linked data.
Keywords
Causal inference
Federated learning
Targeted learning
Projection-based estimand
In this project, we look into high-dimensional Instrumental Variables (IV) models, which have a wide range of applications ranging from econometrics to genomics. Specifically, we focus on the case of many weak IVs and high dimensional exposures and derive a minimax rate for optimal estimators of linear and quadratic functionals in the model. This allow us to infer causal effects and gain information about the signal-to-noise ratio. To obtain a complete picture, we explore versions of the problem in ultra-high dimensional regimes, under suitable notions of sparsity, and proportional asymptotic regimes, without requiring sparsity assumptions. Our analyses rely on a combination of results from causal inference, high dimensional statistics, random matrix theory, and semiparametric inference.
Keywords
High Dimensional Statistics
Instrumental Variables Models
Asymptotic Inference
Causal Inference
Minimax Optimality
Mediation analysis has gained significant attention for its ability to identify causal mechanisms underlying the relationships between a treatment and an outcome via intermediate variables. Despite the development of numerous methodologies, approaches tailored to zero-inflated outcome with explicit causal effects decomposition and interpretation remain underexplored. In this paper, we propose a flexible Bayesian mediation analysis framework capable of handling both zero-inflated count and zero-inflated continuous outcomes, while providing a clear causal effect decomposition. The framework supports a broad range of exposure and mediator distributions. This novel technique employs Bayesian models for both the mediator and the outcome, leveraging Markov Chain Monte Carlo algorithms for parameter estimation. To facilitate implementation, we developed an easy-to-use R package, mediationBayes. Simulation studies demonstrate strong performance in terms of point estimate accuracy and coverage probabilities for both overall and decomposed mediation effects. We further applied our method to microbiome data from the American Gut Project, illustrating its practical utility.
Keywords
Mediation
Bayesian
Zero-inflation
Causal effect decomposition
MCMC
Understanding the mechanisms by which exposure mixtures impact human health is an important topic in environmental health. One of the key challenges lies in the latent nature of information about co-occurring exposures, which complicates the identification of pathways linking mixtures to health outcomes. To address this, we propose a Bayesian framework that integrates mediation analysis with latent factor modeling. This method simultaneously estimates mediation effects associated with both individual exposures and latent exposure mixtures, capturing shared variation within exposure mixtures to enable a comprehensive investigation of their collective and component-specific effects on health outcomes. We evaluate the performance of the proposed framework through simulation studies and assess its applicability using real-world data, illustrating its potential to elucidate meaningful pathways in complex exposure-outcome relationships.
Keywords
Causal mediation analysis
Latent factor modeling
Environmental health
Bayesian inference
Exposure mixtures
Estimating population-level treatment effects is challenging when participation involves sequential self-selection, such as screening compliance followed by treatment decisions. We propose a novel framework that integrates inverse probability weighting within a multistate model to address compound selection biases from these decision stages. Specifically, we first introduce an approach to correct for treatment-selection bias alone, then extend it to jointly correct bias from screening non-compliance and treatment choices. Our two-stage weighting approach accounts for systematic differences at both stages, enabling valid estimation of marginal treatment effects. In addition, compared to standard survival models, the multistate framework captures intermediate health states and competing risks, addressing the non-portability of relative risks and improving generalizability. Simulation studies show that ignoring multi-stage bias or using conventional survival models leads to substantial bias, while our method yields unbiased estimates. Application to the Kerala Oral Cancer Screening Trial illustrates the practical utility of the proposed approach.
Keywords
Multistate models
Inverse probability weighting
Selection bias
Population-level treatment effects
Survival analysis
Causal inference
Co-Author(s)
Qing Pan, George Washington University
Li Cheung, National Cancer Institute
First Author
Guannan Chen, George Washington University
Presenting Author
Guannan Chen, George Washington University
Pricing based on individual customer characteristics is widely used to maximize sellers' revenues. This work studies offline personalized pricing under endogeneity using an instrumental variable approach. Standard instrumental variable methods in causal inference/econometrics either focus on a discrete treatment space or require the exclusion restriction of instruments from having a direct effect on the outcome, which limits their applicability in personalized pricing. We propose a new policy learning method for Personalized pRicing using Invalid iNsTrumental variables (PRINT) for continuous treatment that allow direct effects on the outcome. Relying on the structural models of revenue and price, we establish the identifiability condition of an optimal pricing strategy under endogeneity with the help of invalid instrumental variables. Based on this new identification, which leads to solving conditional moment restrictions with generalized residual functions, we construct an adversarial min-max estimator and learn an optimal pricing strategy. Furthermore, we establish an asymptotic regret bound to find an optimal pricing strategy.
Keywords
causal Inference
pricing
instrumental variable