Practical strategies for handling breakdown of multiple imputation procedures | JoVE Visualize

Practical strategies for handling breakdown of multiple imputation procedures

Cattram D Nguyen^1,2, John B Carlin^3,4, Katherine J Lee^3,4

¹Clinical Epidemiology and Biostatistics Unit, Murdoch Children's Research Institute, The Royal Children's Hospital, Flemington Road, Parkville, Victoria, 3052, Australia. cattram.nguyen@mcri.edu.au.

Emerging Themes in Epidemiology|April 2, 2021

View abstract on PubMed

Summary

Multiple imputation (MI) is crucial for incomplete data but can fail due to numerical issues. This study identifies common causes of MI breakdown and offers practical strategies for robust data handling in statistical software.

Area of Science:

Statistics
Data Science
Biostatistics

Background:

Incomplete data is a common challenge in statistical analysis.
Multiple imputation (MI) is a widely accepted method for addressing missing data.
Failures in MI procedures can arise from numerical instability in imputation algorithms.

Purpose of the Study:

To identify common causes of failure in multiple imputation procedures.
To propose practical strategies for overcoming these failures, particularly within Stata software.
To illustrate these strategies with a real-world case study.

Main Methods:

Analysis of common failure modes in multivariate imputation by chained equations (MICE).
Focus on issues such as perfect prediction and collinearity.
Development and application of troubleshooting strategies.

Main Results:

Identified perfect prediction and collinearity as frequent causes of MI procedural breakdown.
Demonstrated the effectiveness of strategies like imputing composite variables, incorporating prior information, and modifying model forms.
Successfully applied these strategies in a case study using the Longitudinal Study of Australian Children data.

Conclusions:

Numerical problems in MI, especially with MICE, can be overcome with targeted strategies.
Implementing these solutions enhances the reliability of statistical analyses with incomplete datasets.
The proposed methods offer practical guidance for researchers encountering MI failures.