Feature selection methods for optimal design of studies for developmental inquiry

Objectives: As diary, panel, and experience sampling methods become easier to implement, studies of development and aging are adopting more and more intensive study designs. However, if too many measures are included in such designs, interruptions for measurement may constitute a significant burden for participants. We propose the use of feature selection-a data-driven machine learning process-in study design and selection of measures that show the most predictive power in pilot data. Method: We introduce an analytical paradigm based on the feature importance estimation and recursive feature elimination with decision tree ensembles and illustrate its utility using empirical data from the German Socio-Economic Panel (SOEP). Results: We identified a subset of 20 measures from the SOEP data set that maintain much of the ability of the original data set to predict life satisfaction and health across younger, middle, and older age groups. Discussion: Feature selection techniques permit researchers to choose measures that are maximally predictive of relevant outcomes, even when there are interactions or nonlinearities. These techniques facilitate decisions about which measures may be dropped from a study while maintaining efficiency of prediction across groups and reducing costs to the researcher and burden on the participants.