Missing Data in OHCA Registries: How Multiple Imputation Methods Affect Research Conclusions-Paper II.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Missing data in clinical observational studies, like out-of-hospital cardiac arrest (OHCA) registries, can significantly impact statistical validity.
• Complete-case analysis (CCA) and single imputation methods are insufficient as they do not account for imputation uncertainty.
• Multiple imputation (MI) is a standard for missing-at-random (MAR) data but presents implementation challenges.

Purpose of the Study:

• To evaluate the performance of MI in association analysis for OHCA data.
• To compare MI methods against CCA and single imputation techniques.
• To identify optimal MI algorithms for handling missing data in OHCA registries.

Main Methods:

• A simulation framework using Singapore OHCA registry data (N=13,274) with 20%, 30%, and 40% artificially introduced MAR data.
• Implementation of MI using predictive mean matching (PMM), random forest (RF), and classification and regression trees (CART) with 5-20 imputations.
• Assessment of bias and precision in logistic regression models for alert issuance and bystander CPR association.

Main Results:

• CART demonstrated superior performance over PMM, yielding more accurate coefficients and stable confidence intervals (CIs) across missingness levels.
• K-Nearest Neighbors (KNN) provided similar point estimates but underestimated imputation uncertainty.
• PMM exhibited greater bias and less stable CIs, sometimes performing comparably to CCA; MI methods appropriately captured imputation uncertainty with wider CIs than single imputation.

Conclusions:

• The efficacy of MI in OHCA research is highly dependent on the chosen imputation algorithm.
• Tree-based methods, specifically CART and RF, proved most robust and consistent for OHCA data analysis.
• MI utilizing CART/RF is recommended for reliable conclusions in OHCA research, while PMM should be used cautiously.