Variable selection models based on multiple imputation with an application for predicting median effective dose and...

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Missing covariates pose significant challenges for statistical variable selection and prediction.
• Combining results from multiple imputation (MI) for variable selection remains unclear due to potential discrepancies across imputations.
• Existing methods like sparse partial least-squares (SPLS) and elastic net (ENet) have limitations with missing data.

Purpose of the Study:

• To propose a novel MI-based weighted elastic net (MI-WENet) method to handle missing covariate data.
• To develop a robust approach for variable selection and prediction in the presence of missing data.
• To evaluate the performance of MI-WENet against existing methods.

Main Methods:

• The proposed MI-WENet method utilizes stacked MI data with a specific weighting scheme for each observation.
• MI in MI-WENet accounts for sampling and imputation uncertainty.
• A weighting scheme is incorporated to consider observed information within the stacked data.

Main Results:

• Extensive numerical simulations demonstrated the effectiveness of MI-WENet.
• MI-WENet showed superior performance compared to competing alternatives like SPLS and ENet.
• The method was successfully applied to identify predictor variables for endothelial function (ED50 and Emax).

Conclusions:

• The MI-WENet method provides a robust solution for variable selection and prediction with missing covariates.
• This approach effectively integrates multiple imputation and weighted elastic net for enhanced statistical analysis.
• MI-WENet offers a valuable tool for biological and medical research involving complex datasets.