Abstrak
This doctoral dissertation concentrates on data-adaptive prediction of univariate outcomes. It is divided into six chapters. The first chapter is an introduction to loss-based estimation which serves as the foundation for the subsequent chapters. The second chapter introduces the Deletion/Substitution/Addition (D/S/A) algorithm for the prediction of an uncensored outcome with polynomial basis functions. The third chapter explains how the D/S/A algorithm can be used for the prediction of survival times, and it proposes an aggregation scheme for bagging the D/S/A algorithm. The fourth chapter introduces a Super Learner which selects an optimal learning method from various candidate learning methods where the D/S/A algorithm is just one example. The Super Learner is applied to an HIV-1 genotype-phenotype data set which tries to relate mutations in HIV-1 protease and reverse transcriptase to changes in in vitro susceptibility to antiretroviral drugs. The fifth chapter illustrates how to use the D/S/A algorithm for estimation of direct effects which was developed while working with an HIV-1 genotype-clinical outcome data set. Finally, the last chapter is a conclusion and offers suggestions for future work.