Uncovering Underlying Features for State Transition Modeling
Degree granted by The University of Texas at Arlington
Abstract
Modeling of a dynamic system is the representation of the interconnectivity of system state variables and their evolutionary trajectory over time. In this dissertation, the terminology “state transition modeling” refers to a situation when the system state transitions and its evolution is unknown and needs to be estimated. There are situations in many application settings where one does not simply observe the behavior of the system, but also has a desire to take action, intervene, and manipulate one or more system variables, and is interested in seeing the causal effect of the intervention. These interventions within a purely observational setting, as opposed to a randomized controlled trial, have induced a controversial debate between statistics and econometrics. This dissertation presents approaches that seek to uncover the true underlying features for state transition modeling given multiple decisions and set of covariates in a complex situation: finite horizon with a very few stages, non-stationary nature of transitions, and a highly correlated feature space. First, existing state space modeling methodologies are reviewed, and then the proposed methodology is presented in two settings. First, a simplified observational setting is studied with no interventions (i.e., no treatment variables), and then, second, a more complex observational setting is studied with multiple treatment variables under time-varying confounding and multicollinearity.