Abstract
Noncompliance and missing data often occur in randomized trials, which complicate the inference of causal effects. When both noncompliance and missing data are present, previous papers proposed moment and maximum likelihood estimators for binary and normally distributed continuous outcomes under the latent ignorable missing data mechanism. However, the latent ignorable missing data mechanism may be violated in practice, because the missing data mechanism may depend directly on the missing outcome itself. Under noncompliance and an outcome-dependent nonignorable missing data mechanism, previous studies showed the identifiability of complier average causal effect for discrete outcomes. In this paper, we study the semiparametric identifiability and estimation of complier average causal effect in randomized clinical trials with both all-or-none noncompliance and the outcome-dependent nonignorable missing continuous outcomes, and propose a two-step maximum likelihood estimator in order to eliminate the infinite dimensional nuisance parameter. Our method does not need to specify a parametric form for the missing data mechanism. We also evaluate the finite sample property of our method via extensive simulation studies and sensitivity analysis, with an application to a double-blinded psychiatric clinical trial.
Introduction
In randomized controlled trials (RCTs), researchers often encounter two "demons" that threaten the validity of their results: noncompliance and missing data. Noncompliance occurs when participants do not adhere to their assigned treatment group, while missing data arises when participants drop out before the final measurement.
Standard statistical approaches often rely on the "latent ignorability" assumption, which posits that the reason for dropout is independent of the outcome once observed variables are accounted for. However, in many clinical settings—such as psychiatric trials—dropout is often "nonignorable" (Missing Not At Random). For example, a patient might leave a study because their symptoms are worsening, making the missingness directly dependent on the unobserved outcome. This paper introduces a robust semiparametric framework to estimate the Complier Average Causal Effect (CACE) for continuous outcomes under these challenging conditions.
Modeling Causal Relationships and Missingness
To address these issues, it is essential to distinguish between different missing data mechanisms. The relationships between treatment assignment ($Z$), treatment received ($D$), unobserved confounders ($U$), the outcome ($Y$), and the missingness indicator ($R$) can be represented using Directed Acyclic Graphs (DAGs).
In contrast, the authors focus on the "Outcome-Dependent Nonignorable" (ODN) mechanism, where the missingness is a direct consequence of the outcome itself.
The Semiparametric Two-Step Estimator
The core contribution of this work is a Two-Step Maximum Likelihood Estimator (TSMLE). Unlike traditional methods that require researchers to guess the exact mathematical form of the dropout mechanism (e.g., a logistic regression model), this approach is semiparametric. It assumes the outcome follows a distribution within the exponential family (such as Normal, Gamma, or Exponential) but leaves the dropout mechanism completely flexible.
- Step 1: The model estimates the proportions of compliance types—Always-takers, Never-takers, and Compliers—based on the observed treatment assignment and receipt.
- Step 2: It utilizes a conditional likelihood function to estimate outcome parameters. By conditioning the likelihood in a specific way, the unknown dropout function—an "infinite-dimensional nuisance parameter"—mathematically cancels out, allowing for the identification of the CACE without modeling the dropout process directly.
Key Results and Clinical Application
The authors validated their method through extensive simulations and a real-world application:
- Simulation Performance: The TSMLE demonstrated high accuracy and low bias across various distributions. It proved robust even when the ODN assumption was violated, such as when dropout depended on both the outcome and the treatment received.
- Psychiatric Trial Re-analysis: The method was applied to a study comparing Clozapine and Haloperidol for refractory schizophrenia. While standard models assuming ignorable missingness showed a modest benefit for Clozapine, the nonignorable ODN model suggested a more significant clinical effect (a 5.54-point greater reduction in symptom scores). This suggests that traditional models may understate the efficacy of treatments when dropout is outcome-driven.
Conclusion and Future Directions
This paper provides a rigorous mathematical foundation for identifying causal effects in the presence of noncompliance and nonignorable missingness. By bypassing the need to model the dropout mechanism, the TSMLE offers a more robust alternative for clinical trial analysis. Future research will focus on extending this framework to longitudinal data, where patients are tracked over multiple time points, presenting even more complex patterns of missingness.
Summary updated