Flexible variable selection in the presence of missing data.

Area of Science:

• Biostatistics
• Machine Learning
• Bioinformatics

Background:

• Identifying parsimonious feature sets for prediction is crucial but complicated by missing data.
• Existing methods often rely on potentially misspecified finite-dimensional statistical models.
• Misspecification can lead to irrelevant variables and suboptimal predictive panels.

Purpose of the Study:

• To propose a novel nonparametric variable selection algorithm combined with multiple imputation.
• To develop flexible feature panels in the presence of missing-at-random data.
• To achieve control of commonly used error rates and improve classification performance.

Main Methods:

• A nonparametric variable selection algorithm.
• Multiple imputation for handling missing-at-random data.
• Development of strategies for error rate control.

Main Results:

• The proposed method demonstrates good operating characteristics through simulations.
• Achieved higher classification and variable selection performance than penalized regression approaches when models are misspecified.
• Successfully developed biomarker panels for pancreatic cyst classification with complex missingness.

Conclusions:

• The nonparametric approach offers a flexible and robust alternative for variable selection with missing data.
• It outperforms existing methods, particularly when underlying statistical models are misspecified.
• The method has practical utility in biomedical applications, such as biomarker panel development for disease stratification.