How to apply Bayesian stochastic search variable selection with multiply imputed data

Area of Science:

• Psychological research methodology
• Statistical modeling
• Data analysis

Background:

• Modern variable selection methods like lasso and Bayesian approaches are crucial for psychological research to prevent overfitting and enhance model interpretability.
• Addressing missing data is challenging when combined with advanced variable selection techniques, impacting predictor selection accuracy.

Purpose of the Study:

• To demonstrate the implementation of Bayesian variable selection (Stochastic Search Variable Selection - SSVS) with multiply imputed data.
• To compare the effectiveness of SSVS against the least absolute shrinkage and selection operator (lasso) in variable selection with missing data.

Main Methods:

• Utilized stochastic search variable selection (SSVS), a Bayesian method for variable selection.
• Integrated SSVS with multiply imputed datasets to address missing data.
• Employed an imputation and treatment strategy (ITS) for SSVS, compatible with existing software.

Main Results:

• Stochastic search variable selection (SSVS) provides a principled and consistent approach to variable selection with multiply imputed data.
• SSVS demonstrated advantages over lasso in variable selection accuracy in both an example dataset and a simulation study.
• The integration of SSVS with multiply imputed data is straightforward using existing software.

Conclusions:

• Bayesian variable selection, particularly SSVS, offers a robust and reliable method for psychological researchers dealing with missing data.
• SSVS combined with multiple imputation is a superior alternative to lasso for variable selection in the presence of missing data.
• The proposed method is practical and accessible for psychological researchers through existing software implementations.