Recoverability of causal effects under presence of missing data: a longitudinal case study

Area of Science:

• Causal inference
• Missing data analysis
• Pharmacological studies

Background:

• Missing data is a prevalent challenge in longitudinal studies.
• Graphical models offer a framework for handling missing data.
• The CHAPAS-3 trial provides a complex longitudinal dataset for investigation.

Purpose of the Study:

• To assess the utility of graphical models for missing data in a real-world pharmacological study.
• To determine if causal effects can be consistently estimated from available data.
• To explore the impact of missingness mechanisms on estimation accuracy.

Main Methods:

• Utilized missingness-directed acyclic graphs (m-DAGs) to model data gaps.
• Investigated static interventions on multiple continuous variables.
• Proposed and analyzed the 'closed missingness mechanism' concept.
• Conducted simulations and theoretical analyses.

Main Results:

• Recoverability of causal effects is highly sensitive to graph structure.
• Demonstrated consistent estimation is possible under specific m-DAGs.
• Showcased that available case analysis can outperform multiple imputation for missing not at random data.
• Highlighted variability in estimation based on assumed missingness DAGs.

Conclusions:

• Graphical models provide an innovative approach for analyzing complex longitudinal data with missingness.
• The 'closed missingness mechanism' offers a condition for admissible available case analysis.
• Careful consideration of missingness DAGs is crucial for reliable causal effect estimation.