Multiple Imputation Isn’t Always Better: When Complete-Case Regression Works

🧭 Main Claim

Multiple imputation (MI) is often presented as an improvement over listwise deletion (LWD) for regression estimation when data are missing. This paper shows that the complete-case (listwise deletion) estimator can be unbiased even when data are not missing completely at random (MCAR). If the analyst can control for the determinants of missingness, MI offers no advantage over LWD for bias reduction in regression analysis.

🧪 How the argument is presented

The paper provides a clear demonstration and supporting examples that compare the bias, accuracy, and precision of regression estimates under MI and LWD. It highlights when MI is and is not likely to improve inference, and translates those insights into concrete, actionable guidance for researchers.

📌 Key findings

• The complete-case estimator can be unbiased even if data are not MCAR, contrary to common belief.
• When determinants of missingness are controlled in the regression, MI does not reduce bias relative to listwise deletion.
• MI can improve accuracy and precision in some settings, but those settings are specific and identifiable.
• Access to imputation software does not remove the analyst’s responsibility to understand the missing-data process and the data-generating assumptions.

🛠️ Practical guidance offered

• Concrete guidelines are developed to help researchers decide when to use MI versus listwise deletion.
• Recommendations focus on transparency about missingness, controlling for predictors of missingness, and documenting choices that affect inference.

💡 Why it matters

MI remains a useful tool in certain contexts, but it is not a universal fix for missing data problems. Correct application depends on understanding the missing-data mechanism and the variables that determine missingness; following the paper’s guidelines helps increase confidence and transparency in regression results.