Intensive Longitudinal Modeling with Machine Learning Methods

Google map

Abstract: 

Recent advances in artificial intelligence and machine learning (ML) have revolutionized forecasting and classification, even in areas where theoretical understanding is limited. However, in fields such as psychometrics and quantitative psychology, ML has often been applied as a computational tool for fitting relatively structured (e.g., parametric, or semi-parametric) latent variable models. In this talk, I will review and examine issues that arise in using ML methods to analyze intensive longitudinal data under sample size configurations that are realistic or attainable in the social and behavioral sciences. Using simulated data motivated by a laboratory affect study in which ML models are trained with individuals’ physiological data to predict their self-report affect intensity levels, I provide examples to illustrate the strengths and weaknesses of selected ML models designed for intensive longitudinal data, their robustness to changes in hyperparameter values and cross-validation settings, and the interpretability of their outputs in the context of both individual and temporal heterogeneity.