The New England Journal of Statistics in Data Science

Estimating a prediction function is a fundamental component of many data analyses. The super learner ensemble, a particular implementation of stacking, has desirable theoretical properties and has been used successfully in many applications. Dimension reduction can be accomplished by using variable screening algorithms (screeners), including the lasso, within the ensemble prior to fitting other prediction algorithms. However, the performance of a super learner using the lasso for dimension reduction has not been fully explored in cases where the lasso is known to perform poorly. We provide empirical results that suggest that a diverse set of candidate screeners should be used to protect against poor performance of any one screener, similar to the guidance for choosing a library of prediction algorithms for the super learner. These results are further illustrated through the analysis of HIV-1 antibody data.

We consider the problem of developing flexible and parsimonious biomarker combinations for cancer early detection in the presence of variable missingness at random. Motivated by the need to develop biomarker panels in a cross-institute pancreatic cyst biomarker validation study, we propose logic-regression based methods for feature selection and construction of logic rules under a multiple imputation framework. We generate ensemble trees for classification decision, and further select a single decision tree for simplicity and interpretability. We demonstrate superior performance of the proposed methods compared to alternative methods based on complete-case data or single imputation. The methods are applied to the pancreatic cyst data to estimate biomarker panels for pancreatic cysts subtype classification and malignant potential prediction.