Marginal Structural Models for Time-varying Endogenous Treatments: A Time-Varying Instrumental Variable Approach

Eric J. Tchetgen Tchetgen, Haben Michael, Yifan CuiarXiv 2018 · DOI · arXiv

Abstract

Robins (1998) introduced marginal structural models (MSMs), a general class of counterfactual models for the joint effects of time-varying treatment regimes in complex longitudinal studies subject to time-varying confounding. He established identification of MSM parameters under a sequential randomization assumption (SRA), which essentially rules out unmeasured confounding of treatment assignment over time. In this technical report, we consider sufficient conditions for identification of MSM parameters with the aid of a time-varying instrumental variable, when sequential randomization fails to hold due to unmeasured confounding. Our identification conditions essentially require that no unobserved confounder predicts compliance type for the time-varying treatment, the longitudinal generalization of the identifying condition of Wang and Tchetgen Tchetgen (2018). Under this assumption, we derive a large class of semiparametric estimators that extends standard inverse-probability weighting (IPW), the most popular approach for estimating MSMs under SRA, by incorporating the time-varying IV through a modified set of weights. The set of influence functions for MSM parameters is derived under a semiparametric model with sole restriction on observed data distribution given by the MSM, and is shown to provide a rich class of multiply robust estimators, including a local semiparametric efficient estimator.

Introduction

In longitudinal research, estimating the causal effect of a treatment that changes over time is a significant challenge. Researchers typically rely on Marginal Structural Models (MSMs) to account for time-varying confounding. However, the validity of these models traditionally rests on the Sequential Randomization Assumption (SRA). SRA posits that all variables influencing both treatment assignment and the outcome have been measured at every time step.

In real-world clinical or social science settings, this assumption is frequently violated. Unmeasured confounders—such as a patient's underlying health motivation, genetic factors, or unrecorded environmental influences—often create "endogeneity," where the treatment assignment is correlated with the error term of the outcome. When SRA fails, standard estimation techniques like Inverse Probability Weighting (IPW) yield biased and unreliable results. This paper introduces a novel framework to address this "time-varying endogeneity" using instrumental variables (IVs).

The Time-Varying Instrumental Variable Approach

To overcome the limitations of SRA, this research leverages time-varying instrumental variables. An instrument is a variable that influences the treatment but has no direct effect on the outcome except through that treatment. While IV methods are common in cross-sectional studies, applying them to longitudinal MSMs introduces significant complexity.

The Independent Compliance Type Assumption

The core innovation of this paper is the "Independent Compliance Type Assumption." This assumption serves as the longitudinal generalization of recent breakthroughs in cross-sectional IV analysis. It requires that unobserved confounders do not predict a subject's "compliance type"—essentially how they respond to the instrument—over the course of the study. By making this assumption, the authors provide a mathematical pathway to identify causal effects even when "ghost" variables (unmeasured confounders) are present.

From IPW to Modified Weights

Standard MSMs use IPW to create a pseudo-population where treatment is independent of measured confounders. This paper extends this logic by deriving a new set of weights that incorporate the IV. These modified weights effectively "filter out" the bias introduced by unmeasured confounding by modeling the relationship between the instrument and the treatment, rather than just the treatment assignment itself.

Statistical Properties and Robustness

A primary contribution of this work is the derivation of a class of "multiply robust" estimators. In complex longitudinal models, researchers must specify several sub-models: the treatment process, the instrument process, and the outcome process.

Triple Robustness

The authors demonstrate that their proposed estimator, $\hat{\beta}_{mr}$, possesses a "triple robustness" property. This means the estimator remains consistent and asymptotically normal (CAN) if at least one of several combinations of the underlying models is correctly specified. This provides researchers with multiple "safety nets," ensuring that even if parts of the complex longitudinal system are slightly misspecified, the resulting causal estimates remain valid.

Semiparametric Efficiency

Beyond robustness, the paper addresses the issue of statistical precision. By deriving the "efficient influence function," the authors identify a specific estimator within their broad class that achieves the semiparametric efficiency bound. This ensures that the estimator makes the most statistically efficient use of the available data, providing the narrowest possible confidence intervals for the causal effect.

Methodology and Scope

As a technical report, this paper focuses on the rigorous mathematical proofs and the derivation of the statistical theory required to support these new estimators. The derivations carefully respect temporal ordering, ensuring that potential outcomes at any given time point are only influenced by past treatments and instruments, adhering to the "future-cannot-cause-past" constraint inherent in longitudinal data.

Currently, the framework is optimized for binary (yes/no) treatments at each time step. The authors acknowledge that extending this to polytomous or continuous treatments is a subject for future research. Additionally, while the theoretical foundations are established, the authors note that these multiply robust estimators are computationally intensive compared to standard IPW equations.

Conclusion

This paper bridges a critical gap between the Marginal Structural Models used in epidemiology and the Instrumental Variable methods favored in economics. By providing a robust toolkit for handling time-varying endogeneity, it offers a path forward for causal inference in longitudinal studies where unmeasured confounding was previously an insurmountable barrier. While empirical validation and software implementation are currently underway, the theoretical framework established here provides the necessary foundation for the next generation of robust longitudinal analysis.

Summary updated

Abstract

Robins (1998) introduced marginal structural models (MSMs), a general class of counterfactual models for the joint effects of time-varying treatment regimes in complex longitudinal studies subject to time-varying confounding. He established identification of MSM parameters under a sequential randomization assumption (SRA), which essentially rules out unmeasured confounding of treatment assignment over time. In this technical report, we consider sufficient conditions for identification of MSM parameters with the aid of a time-varying instrumental variable, when sequential randomization fails to hold due to unmeasured confounding. Our identification conditions essentially require that no unobserved confounder predicts compliance type for the time-varying treatment, the longitudinal generalization of the identifying condition of Wang and Tchetgen Tchetgen (2018). Under this assumption, We derive a large class of semiparametric estimators that extends standard inverse-probability weighting (IPW), the most popular approach for estimating MSMs under SRA, by incorporating the time-varying IV through a modified set of weights. The set of influence functions for MSM parameters is derived under a semiparametric model with sole restriction on observed data distribution given by the MSM, and is shown to provide a rich class of multiply robust estimators, including a local semiparametric efficient estimator.
BibTeX
@misc{tchetgen2018marginal,
  author = {Eric J. Tchetgen Tchetgen and Haben Michael and Yifan Cui},
  title = {{Marginal Structural Models for Time-varying Endogenous Treatments: A Time-Varying Instrumental Variable Approach}},
  year = {2018},
  eprint = {1809.05422},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1809.05422},
}
Figure