Multiple imputation in the presence of high-dimensional data

Area of Science:

• Statistics
• Biomedical Research
• Epidemiology

Background:

• Missing data is common in research, potentially causing bias and reduced efficiency.
• Multiple imputation is a popular method for handling missing data.
• Optimal multiple imputation strategies for high-dimensional data remain unclear.

Purpose of the Study:

• To investigate and compare multiple imputation methods for high-dimensional data.
• To evaluate regularized regression and Bayesian lasso regression for imputing missing values.
• To determine the most effective imputation strategy in high-dimensional settings.

Main Methods:

• Numerical studies were conducted to compare imputation methods.
• Investigated approaches using regularized regression and Bayesian lasso regression.
• Evaluated the impact of data dimension, active set size, and correlation strength.

Main Results:

• Standard multiple imputation performed poorly with high-dimensional data.
• Bayesian lasso regression generally outperformed other imputation methods.
• Bayesian lasso showed better performance than standard imputation even with a correctly specified model.

Conclusions:

• Bayesian lasso regression is well-suited for multiple imputation in high-dimensional data.
• Extensions of Bayesian lasso may offer further improvements.
• This method provides a more robust approach compared to standard regression techniques.