|
Abstract:
|
In establishing prediction models, much effort is centered around better identifying predictors (e.g., aided by machine learning methods). This paper focuses on improving the outcome side of the models - a rather neglected aspect of prediction model development. We are particularly interested in using longitudinal information to optimally characterizing and classifying individuals' outcome status, and validating the formulated prediction targets. We explore the joint use of empirical model fitting, clinical insights, and cross-validation based on how well formulated targets are predicted by clinically relevant antecedents to triangulate valid prediction targets. To this end, we expand the use of latent trajectory types as predictors of future outcomes and as concurrent outcomes that develop in parallel with other relevant outcomes. The proposed approach is illustrated using data from the Longitudinal Assessment of Manic Symptoms study, where longitudinal development types of manic symptoms among children are modeled as outcomes that are closely related to other psychiatric disorders (e.g., depression), and distal outcomes (e.g., risky health behaviors in adulthood).
|
