Delineating Discrete Generalizable States from Intraindividual Time Series: Towards a Science of Moments

The present paper pursues archetypal states—compound sets of concurrent, fixed-distance intervals in temporal variables that predict person-level data across persons. Using unsupervised learning, we identify a set of states defined by varying degrees of negatively correlated positive and negative affect. We demonstrate the consistency of these structures across three samples. Sample 1 (N=838) was split into N=500 training series and N=338 hold-out series. Training data were used to distill archetypal compound emotion states, which were validated across the hold-out sample and two external samples—a naturalistic sample of 179 participants and a sample of 45 individuals with depression and anxiety. Predictions of momentary variation in the out-of-sample data accounted for 40% to 50% of the variance in these unseen data. We propose that the current paradigm serves as a proof of concept for a novel and generative science of moments that provides means for transcending the idiographic-nomothetic divide.