Background: Chronic disorders are among today’s most pressing public health issues. Important components of chronic care treatment are the dynamic treatment strategies that offer a way to evaluate the sequential decision making involved in adaptive clinical practice and thereby offer a potential way to improve patient management. Formally, a dynamic treatment strategy is an algorithm that dictates how to treat a patient over time, taking into account both treatment history and clinical response to those treatments. Since it is unlikely that there is access to large random samples of patients treated under each time-varying strategy, decisions need to be based upon analyzing complex longitudinal data from observational studies.
Proposed Solution to the Problem: Conventional statistical methods, however, cannot appropriately adjust for confounding found in dynamic treatments. Furthermore, in contrast to g-methods including inverse-probability weighting (IPW) and the parametric g-formula, TMLE and machine leaning techniques are cutting-edge tools to compare dynamic strategies using longitudinal observational data. There is a need to advance dialogue among the research community on how to apply and implement these advanced CER methods for comparison of dynamic treatments.
Objectives: Motivated by the increasing availability of complex longitudinal databases and the growing interest in using advanced causal inference methods for CER, the project team proposes to conduct three two-hour online workshops in 2019 regarding use of causal inference techniques to address real-world time-varying treatment questions.
Activities: Convening three workshops as follows.
- Workshop 1: Defining a Causal Question, focuses on how to formulate a well-defined causal question regarding time-varying treatments and how to frame study designs to effectively leverage electronic health data
- Workshop 2: Applying TMLE and Machine Learning Techniques
- Workshop 3: Applying G-Methods, focuses on the use of the IPW and the parametric g-formula method for CER. Survey evaluations by attendees will further refine workshop objectives and delivery of information.
Outcomes and Outputs (projected): By promoting broader use of advanced statistical methods, the team anticipates this project will facilitate the provision of scientifically valid information needed to help patients and providers make informed decisions.
Patient/Stakeholder Engagement Plan: These workshops are intended for applied clinical researchers, stakeholders interested in research, statisticians, epidemiologists, and health services researchers who are interested in using causal models to address complex CER questions and already have a good understanding of CER methods for fixed interventions. The project team does not anticipate inviting non-researcher clinicians or those not familiar with the fundamentals of causal inference methods.
Project Collaborators: Maya Petersen, MD, PhD, Associate Professor of Biostatistics and Epidemiology, whose methodological research focuses on using novel causal inference methods to problems in health, with an emphasis on longitudinal data and dynamic regimes and machine learning methods; Jessica Young, PhD, Associate Professor of Biostatistics, whose research has focused on the theoretical development of parametric g-formula and IPW methods to address the complications inherent in real-world observational studies using observational data with time-varying confounding.