Keywords: Synthetic Control Arm (SCA), Propensity score, Historical controls, External controls
Room: Wilson A
Propensity score (PS) techniques are a popular tool used in observational studies to aid in estimation of treatment effects by minimizing the effects of observed confounding factors. However, PS techniques are rarely used in the field of clinical trials where randomized controlled trials (RCTs) are the gold standard commonly implemented for the confirmatory pivotal trials and any potential confounders are assumed balanced between the treatment groups due to the randomization process. With an increasing amount of clinical trial data available and wide use of electronic data collection systems in clinical trials, there is great appeal for patients and researchers in the possibility of gaining efficiencies in future clinical trials by using historical clinical trials data, especially in diseases with ethical or practical challenges. In addition, early availability of investigational product on the market through FDA’s mechanisms of breakthrough therapy designation as well as fast track, accelerated approval and priority review may challenge the confirmatory trial with slow enrollment, early drop-out and on-study treatment cross-over. A Synthetic Control Arm (SCA), a well-matched external control made from patient-level data from previous clinical trials, has been proposed to augment or replace a randomized control in such settings. Various PS-based methods, such as matching or weighting, are proposed in building SCA by appropriately selecting the historical patients to balance or properly weight the composition of the SCA in terms of baseline characteristics with that of the investigational group. SCA may offer a valuable option for reducing recruitment, retention, and compliance concerns without compromising the scientific understanding of the treatment effect in clinical studies. Is propensity score methodologies an efficient tool to build SCA? Is it crucial to examine the impact of unmeasured confounders? Are there statistical methods to assess the impact?