Targeting Clinically Relevant Causal Estimands in Trials With Latent Exposure-Induced Intermediate Confounding Through Structural Equation Modeling and Propensity Score Weighting

Abstract

Introduction

Evaluating therapeutic efficacy in heart failure with reduced ejection fraction (HFrEF) is often challenging because improvements in the primary outcome, such as left ventricular ejection fraction (LVEF), may be influenced by complex mechanistic pathways. A methodological issue in HFrEF trials arises when congestion acts as an exposure-induced intermediate confounder of the relationship between rescue therapy, a post-randomization intercurrent event, and the primary outcome. Separating the true treatment effect from the impact of clinical rescue is therefore essential to understand the therapy’s underlying mechanistic pathways. This objective is closely aligned with the ICH E9(R1) addendum, which emphasizes the need to translate clinical questions into clearly defined causal estimands in order to appropriately address intercurrent events. Moreover, since congestion is not directly observed as a single variable, it may be conceptualized as a latent variable measured via correlated indicators subject to measurement error.

Objectives

The primary objective of this study is to evaluate the performance of Structural Equation Modeling (SEM) in estimating the Controlled Direct Effect (CDE) in the presence of latent intermediate confounding and measurement error. We identify the Marginal Structural Model (MSM) as a standard causal inference benchmark utilizing Inverse Probability Weighting (IPW) to compare its performance against the SEM in terms of bias, efficiency, and statistical power.

Methods

We employed a simulation study to assess estimator performance across a range of sample sizes (n = 100 to 1000) and target CDE scenarios. The data-generating mechanism was designed to either satisfy or violate the Positivity Assumption. The CDE was defined as the effect of heart failure therapy on LVEF not operating through rescue therapy (M): CDE_M = E[Y_1,m] − E[Y_0,m], where M is fixed to the same value under both treatment levels. The comparative framework contrasted SEM, which explicitly incorporates latent variables and measurement error for indicators of congestion, against MSM. Performance metrics included Mean Bias, RMSE, Statistical Power, and Type I Error Rate.

Results

Analysis of estimator bias revealed that SEM demonstrated superior performance, maintaining significantly lower mean bias than MSM weighting approaches, as the strength of the intermediate confounding effect increased. In terms of statistical power, SEM showed increasing performance as the sample size increased, reaching the target 80% power threshold and properly identifying the CDE across all simulated effect sizes. SEM successfully avoided practical positivity violations observed in MSM. While MSM struggled with the high dimensionality of the intermediate confounders, SEM employed latent-variable dimension reduction to maintain estimation stability.

Conclusions

SEM provides a robust and flexible unified framework for estimating clinically meaningful causal effects in complex HFrEF trials. Our findings suggest that, when combined with a well-specified causal graph and explicit identification assumptions, SEM is a powerful tool for pre-specified causal estimands rather than merely an exploratory path-analysis tool. SEM is highly effective for addressing intercurrent events, such as rescue therapy, in compliance with ICH E9(R1) guidelines. Treatment–mediator interaction will be addressed in future studies.