From Path Coefficients to Targeted Estimands: A Comparison of Structural Equation Models (SEM) and Targeted Maximum Likelihood Estimation (TMLE)

2025-11-02

AI Summary

This study addresses the bias in causal inference arising from model misspecification in structural equation modeling (SEM). It compares SEM’s performance against targeted maximum likelihood estimation (TMLE) in terms of robustness. We propose a doubly robust framework integrating nonparametric structural modeling, TMLE, and machine learning—thereby relaxing SEM’s stringent parametric assumptions. Simulation results demonstrate that TMLE substantially outperforms SEM in bias reduction, mean squared error, and confidence interval coverage. An empirical application examining the effect of poverty on high school enrollment further corroborates these findings: the direct effect identified by SEM becomes statistically insignificant under TMLE, highlighting TMLE’s resilience to misspecification of confounding and mediation structures. This work advances causal mediation analysis by offering a more reliable, assumption-lean estimation paradigm grounded in semiparametric efficiency theory and robust statistical learning.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Structural Equation Modeling (SEM) has gained popularity in the social sciences and causal inference due to its flexibility in modeling complex relationships between variables and its availability in modern statistical software. To move beyond the parametric assumptions of SEM, this paper reviews targeted maximum likelihood estimation (TMLE), a doubly robust, machine learning-based approach that builds on nonparametric SEM. We demonstrate that both TMLE and SEM can be used to estimate standard causal effects and show that TMLE is robust to model misspecification. We conducted simulation studies under both correct and misspecified model conditions, implementing SEM and TMLE to estimate these causal effects. The simulations confirm that TMLE consistently outperforms SEM under misspecification in terms of bias, mean squared error, and the validity of confidence intervals. We applied both approaches to a real-world dataset to analyze the mediation effects of poverty on access to high school, revealing that the direct effect is no longer significant under TMLE, whereas SEM indicates significance. We conclude with practical guidance on using SEM and TMLE in light of recent developments in targeted learning for causal inference.

Problem

• Comparing SEM and TMLE for causal effect estimation
• Evaluating robustness to model misspecification in causal methods
• Analyzing mediation effects of poverty on education access

Innovation

• TMLE uses doubly robust machine learning
• TMLE outperforms SEM under model misspecification
• TMLE builds on nonparametric structural equation models