Dynamic Treatment Regimes for Within- and Between-Group Interference in Clustered and Hierarchical Datasets

Abstract

Precision medicine is an interdisciplinary field that aims to tailor treatments based on an individual’s unique characteristics. Dynamic treatment regimes (DTRs) formalize this process through step-by-step decision rules that utilize patient-specific information at each stage of analysis and subsequently output recommendations for the optimal course of action. To date, much of the biostatistical literature has analyzed DTR estimation under the assumption of no interference; that is, a given individual’s outcome is not affected by the treatments received by others. However, this assumption is often violated in a variety of social and spatial networks, such as households and communities, and particularly in the context of infectious diseases or resource allocation. Although some recent developments have been made for DTR estimation for couples in households and in networks where general forms of interference are taking place, it has yet to be shown how DTRs can be estimated for individuals in clustered networks where interference may occur within and between predefined groups. Specifically, our attention shifts toward hierarchical networks, which provide a framework where interference can occur both within and between groups in the same hierarchy but not across hierarchies.

This thesis contains three main projects that contribute to DTR estimation in the context of service utilization within the healthcare system and under interference networks. In Chapter 3, we show how the dynamic weighted ordinary least squares regression (dWOLS) DTR methodology can be applied to determine whether a patient would benefit from being discharged to home or admitted to a hospital using data simulated from a retrospective cohort study for patients who experienced an opioid-related overdose in British Columbia. This project was inspired from collaboration with a provincial health authority in Canada, which has resulted in the draft of a manuscript that can be submitted for publication following an ethics approval. In Chapter 4, we demonstrate an application of dWOLS to hierarchical networks where both within- and between-group interference takes place, and through a series of simulation studies, we show how incorrectly assuming that there only exists within-group interference can result in lower rates of optimal treatments assigned, particularly as the number of subgroups (and thus the potential for between-group interference) increases. Our models assess the performance of “standard” overlap weights constructed from logistic mixed models with nested random intercepts as well as network weights that are constructed from a joint propensity score. We apply our findings to the simulated opioid overdose dataset, where our goal is to illustrate how sequences of recommended dispositions for opioid overdose patients can be optimized if interference occurs within and between different facilities. Finally, in Chapter 5, we propose several sets of simulation studies to assess the extent to which interference may be taking place between individuals to warrant use of modified dWOLS methodology that accounts for interference.