Recoverability and estimation of causal effects under typical multivariable missingness mechanisms

Area of Science:

• Causal inference
• Missing data methods
• Epidemiology

Background:

• Identifiability of average causal effect (ACE) relies on causal and missingness assumptions.
• Missingness directed acyclic graphs (m-DAGs) visualize missingness mechanisms.
• Prior work on ACE recoverability excluded effect modification and estimation methods.

Purpose of the Study:

• Extend research on ACE recoverability to settings with effect modification.
• Evaluate performance of missing data methods for ACE estimation using g-computation.
• Investigate ACE estimation in the context of canonical missingness mechanisms.

Main Methods:

• Determined ACE recoverability in settings with effect modification.
• Conducted simulation study evaluating complete case analysis (CCA) and multiple imputation (MI).
• Used correctly specified g-computation for ACE estimation in simulations based on the Victorian Adolescent Health Cohort Study (VAHCS).

Main Results:

• ACE is recoverable if no variable (exposure, outcome, confounder) causes its own missingness, excluding unmeasured confounders of missingness indicators.
• Multiple imputation (MI) methods compatible with g-computation yielded approximately unbiased ACE estimates across most canonical m-DAGs.
• Exceptions for MI include when the outcome causes its own missingness or the missingness of a variable that causes its own missingness.

Conclusions:

• ACE recoverability is contingent on specific missingness assumptions within m-DAGs.
• Compatible MI methods offer robust ACE estimation, but sensitivity analyses may be needed in challenging missingness scenarios.
• Findings have practical implications for causal effect estimation in observational studies with missing data.