When Is a Complete-Case Approach to Missing Data Valid? The Importance of Effect-Measure Modification.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Missing data handling is crucial for unbiased causal effect estimation in epidemiology.
• Complete-case analysis (CCA) is a common method, but its validity depends on missing data mechanisms.
• Previous research indicates CCA yields unbiased stratified effect estimates when missingness is independent of the outcome, conditional on exposure and covariates.

Purpose of the Study:

• To assess the bias introduced by CCA for adjusted marginal causal effects under confounding.
• To investigate bias under various causal structures of missing data.

Main Methods:

• The study evaluates bias in marginal effect estimation using CCA.
• It considers different causal structures for missing data and the presence of confounding.
• The analysis focuses on the requirements for an unbiased estimate of the unconditional joint distribution of confounders and covariates.

Main Results:

• Estimating the marginal risk difference requires an unbiased joint distribution of confounders and covariates influencing missingness.
• The dependence of missing data on these covariates must be accounted for to ensure valid covariate distribution estimates.
• CCA is unbiased for marginal risk differences if stratified effect estimates are unbiased and no effect-measure modification exists on the absolute scale.

Conclusions:

• CCA can be biased for adjusted marginal effects when confounding is present and missing data depend on covariates.
• Valid estimation of marginal effects using CCA necessitates careful consideration of the causal structure and potential effect-measure modification.
• Understanding the interplay between missing data, confounding, and effect modification is key for accurate causal inference.