Variable Selection and Subsequent Case Prediction in Semi-parametric Models

Prediction of health status is a novel technique of forecasting the future health conditions with existing knowledge and available data. A reliable statistical model can lead to high performance of health status prediction. Built upon a semi-parametric variable selection approach, an algorithm to predict health conditions is developed and assessed. This algorithm is compared with three competing prediction methods based on logistic regressions, random forest, and support vector machines. Four statistics, accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC), are used to compare the performance across different approaches. The proposed approach, based on the simulation findings, does not perform as expected with respect to its ability to handle complex joint effects. The methods are then implemented and compared on the prediction of asthma status at the age of 10 years based on variables selected from 45 candidate variables available in the Isle of Wight birth cohort