Advances in mobile technology have enabled scientists to study behavior in the individual's natural environment. In these settings, mechanistic theory is based on latent constructs. In a behavioral mHealth study, multimodal data is collected both passively and actively via mobile devices and sensors. Often these measurements are directly linked to the latent constructs coming out of behavioral theory. In mHealth, a critical composite, dynamic, hypothetical latent construct is engagement. A joint model for the dynamic health behavior and engagement outcomes is required to inform their dynamic interrelationship. I will introduce a hierarchical, partially observable Gaussian process model that depends on dynamic exogenous variables and multiple noisy measures of the latent construct. The model accounts for interventions that can lead to increased step counts and decreased engagement as well as alter their correlation. We present analysis of HeartSteps, a mobile health intervention study aimed at increasing physical activity.