Design and analysis approaches with focus on long-term observational studies in oncology’
View Presentation View Presentation
*Larry Leon, Genentech 

Keywords: Observational study design, Registry study design, Oncology, Time-dependent confounding

It is well known that there are many sources of potential biases when conducting and analyzing observational studies. Retrospective analyses can be fraught with additional unintended biases due to internal influences from the sponsor. Unfortunately, it is not uncommon for observational studies to be non-specific in their study design objectives and thus analyses are mostly retrospective. We suggest that the aforementioned type of biases can be minimized by prospectively specifying the specific questions and analysis methods at the study design stage. We discuss our experiences with a long-term registry study of Lung cancer (NSCLC) patients and propose approaches to avoid common pitfalls when comparing therapies in Oncology settings.

In particular, comparing the effectiveness of treatment regimens in long-term observational studies wherein patients’ therapies may vary over time is challenging. For example, in oncology clinical practice ‘whether’, ‘when’, and for ‘how long’ to treat is a dynamic process involving many factors, including potential time-dependent confounding factors. We discuss some modeling approaches, and evaluate their performance in a simulation experiment designed to mimic a study with time-varying treatments in the presence of time-dependent confounding.