Variable Selection with Multiply Imputed Data When Considering Interaction Effects

• Activity Number: 366
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Aya Mitani*+ and Allison W. Kurian and Amar K. Das and Manisha Desai
• Companies: Stanford University and Stanford University and Dartmouth University and Stanford University
• Keywords: Variable selection ; Missing data ; Multiple imputation ; Interaction effects

Abstract

Missing data is an unavoidable problem in medical and epidemiological studies. Multiple imputation (MI) has been gaining popularity as a theoretically sound and practical method for handling missing data. Variable selection techniques are often employed for the purpose of building parsimonious models, doing prediction, or identifying relevant risk factors for an outcome. Incorporating MI into the variable selection process, however, is not straightforward and is particularly complicated when synergistic effects are considered. Through simulations, we compare the properties of parsimonious models developed from various approaches of both variable selection techniques and applications of MI, such as whether to include all possible interaction effects initially in the imputation model or to only consider main effects with subsequent testing of interaction effects in the model building process. Scenarios with and without the presence of an interaction effect are examined. We provide practical guidelines for implementing MI when employing variable selection, and illustrate differences among the proposed methods using real data from a breast cancer study.