TL;DR
This paper tackles the challenge of estimating causal effects in clinical trials where patients don't always follow treatment instructions (noncompliance) and some outcome data is missing in a way that might be related to the outcome itself (nonignorable missing data). They propose a new statistical method that can handle both issues simultaneously for continuous outcomes, offering a more robust way to understand treatment effectiveness when data is imperfect.
What Problem Does This Paper Address?
In clinical trials, the gold standard for determining if a treatment works is randomization. However, real-world trials often face two major hurdles:
- Noncompliance: Participants might not take the assigned treatment as instructed. This means the treatment they actually received might differ from what they were assigned.
- Missing Data: Not all participants have their outcomes measured. If the reason data is missing is related to the outcome itself (e.g., sicker patients are more likely to drop out), this is called "nonignorable missing data" and can lead to biased results if not handled properly.
When both noncompliance and nonignorable missing data are present, estimating the true causal effect of a treatment becomes very difficult. Previous methods often assumed that missing data was "ignorable" (meaning it wasn't related to the outcome) or focused only on binary outcomes. This paper addresses the gap for continuous outcomes where missingness is nonignorable and noncompliance exists, specifically focusing on the "Complier Average Causal Effect" (CACE) – the effect of the treatment on those who would comply with their assigned treatment.
Methods
The authors propose a novel two-step maximum likelihood estimator (TSMLE). Here's a breakdown of their approach:
- Principal Stratification: They frame the problem within the principal stratification framework, which categorizes individuals based on their potential compliance behavior (e.g., "always-takers," "compliers," "defiers"). This helps to isolate the group of interest: compliers.
- Outcome-Dependent Nonignorable (ODN) Missing Data Mechanism: They explicitly model the scenario where the probability of an outcome being missing depends on the outcome itself. This is a more realistic assumption than "missing at random" in many clinical settings.
- Semiparametric Approach: Crucially, their method doesn't require specifying a parametric form for the missing data mechanism. Instead, they assume the outcome distribution belongs to the exponential family. This offers flexibility and reduces the burden of strong distributional assumptions.
- Two-Step Estimation:
- Step 1: They first estimate parameters related to treatment assignment and compliance using maximum likelihood estimation.
- Step 2: In the second step, they use a conditional likelihood approach to estimate the parameters of the outcome distribution. This step is designed to be independent of the unknown missing data mechanism, thus avoiding the need to model it parametrically.
This two-step process allows them to disentangle the effects of noncompliance, missing data, and the treatment itself, leading to a more reliable estimate of the CACE.
Datasets
The paper uses a real-world dataset from a double-blinded randomized clinical trial comparing clozapine and haloperidol for refractory schizophrenia. This dataset has been previously analyzed in other studies, allowing for comparison.
Key Insights and Ideas
- ODN is More Realistic: The paper highlights that assuming missing data is "ignorable" (LI) is often too simplistic in clinical trials. The outcome-dependent nonignorable (ODN) assumption, where missingness depends on the outcome, is more plausible and can lead to different, and potentially more accurate, conclusions.
- Semiparametric Power: By only assuming the outcome distribution is from the exponential family and not the missing data mechanism, the method gains robustness. This is a significant advantage over methods that require fully parametric models for both.
- Identifiability is Key: A core part of the paper is demonstrating that the CACE is identifiable under their proposed assumptions. This means that, in principle, the true causal effect can be uniquely determined from the observed data, even with noncompliance and nonignorable missingness.
- Two-Step Method's Robustness: The simulation studies suggest that their proposed method is robust to violations of some assumptions, performing well across different outcome distributions and missing data scenarios.
Key Results
- Identifiability Theorem: The paper proves that the Complier Average Causal Effect (CACE) is semiparametrically identifiable under their set of assumptions, including randomization, monotonicity, compound exclusion restrictions, and the ODN missing data mechanism.
- TSMLE Performance: Simulation studies with 10,000 replications show that their TSMLE method yields small biases and standard deviations for CACE estimates across various distributions (normal, exponential, gamma, log-normal). The empirical coverage proportions of their confidence intervals are close to the nominal 95%.
- Comparison with LI: When comparing their ODN-based method with a standard Likelihood-based method (LI) that assumes ignorable missing data, their method generally performs better, especially when the ODN assumption is violated by the LI method. The LI method shows larger biases and poorer coverage when the missing data mechanism is truly outcome-dependent.
- Application to Psychiatric Trial: When applied to the clozapine vs. haloperidol trial, their method provides estimates for CACE. While some results were insignificant (95% CI including zero), the 90% confidence intervals showed differences depending on whether homoskedastic or heteroskedastic assumptions were made for the outcome distribution, suggesting the importance of considering outcome variability.
Limitations and Future Work
- Exponential Family Assumption: While more flexible than fully parametric models, the assumption that the outcome distribution belongs to the exponential family is still a restriction. The paper notes that for some distributions (e.g., when K > 3 in their formulation), identifiability might be an issue, and they provide a counterexample.
- Cross-Sectional Data: The current work focuses on cross-sectional data. The authors suggest that generalizing their methods to longitudinal data (where outcomes are measured over time) is a promising area for future research.
- More Complex Missing Mechanisms: While ODN is more realistic than LI, real-world missingness can be even more complex, potentially depending on treatment assignment (Z), compliance (U), and received treatment (D) in intricate ways. Exploring these more general missing data mechanisms could be a future direction.
Summary updated