TABLE 3.

Candidate algorithms included in the ensemble Super Learner library

Prediction Algorithm	(approximate) Controls per Case	Eligible Covariates	Notes
Logistic Regression	50, 20, or 10	All	Weighted
(14 variants)	50, 20, or 10	Pre-Selected	Weighted
	50, 20, or 10	All
	50, 20, or 10	Pre-selected
	50	All	Backward selection
	50	Pre-Selected	Backward selection
Lasso	20 or 10	All	Deviance loss
(8 variants)	20 or 10	All	neg AUC loss
	20 or 10	Pre-Selected	Deviance loss
	20 or 10	Pre-Selected	neg AUC loss
Ridge Regression	20 or 10	All	Deviance loss
(8 variants)	20 or 10	All	neg AUC loss
	20 or 10	Pre-Selected	Deviance loss
	20 or 10	Pre-Selected	neg AUC loss
Random Forest	50, 20, or 10	All	10,000 trees, 1/3 of covariates sampled per split
(6 variants)	50, 20, or 10	Pre-Selected	10,000 trees, 1/3 of covariates sampled per split
Support Vector Machine	50	All	Tuning parameters chosen by cross validation
(2 variants)	50	Pre-Selected	Tuning parameters chosen by cross validation
Neural Network	20 or 10	Pre-Selected	10 nodes in 1hidden layer
(4 variants)	20 or 10	Pre-Selected	5 nodes in 1 hidden layer

Note: Risk scores were re-scaled to account for undersampling of controls, except for weighted logistic regression algorithms.