General Considerations of RWE/RWD

Organizers: Qi Jiang (Seagen), Jean Pan (Amgen), Yeh-Fong Chen (FDA)
Chair: Biao Xing (Seagen)
Vice Chair: Qi Jiang (Seagen)

Speakers:
Wei Hua (FDA)
Ram Tiwari (BMS)
Andy Wilson (PAREXEL)

Abstracts:

Title: Epidemiological considerations of real-world evidence
Speaker: Wei Hua (FDA)

Under the 21st Century Cures Act, real-world evidence (RWE) is considered for regulatory decision making. In contrast to traditional randomized and controlled trials, there are various challenges for epidemiological studies using real-world data (RWD) to meet the requirement for substantial evidence. This presentation will briefly discuss RWD/RWE as defined in the Framework for FDA’s Real-World Evidence Program, substantial evidence for drug approval per 21 CFR 314.126, and common challenges and considerations from an epidemiological perspective.

Title: Leveraging external evidence for augmenting a single-arm study
Speaker: Ram Tiwari (BMS)

There is a wide variety of study designs that involve the leveraging of external data. These external data can be leveraged to augment a single-arm study, or construct or augment either the control arm or the treatment arm (or both) of a comparative clinical study, and they may come from a single source or from multiple sources. In this presentation, we will briefly discuss the Medical Device Innovation Consortium (MDIC) External Evidence (EEM) Framework and then present a propensity-score based method for leveraging an external data to augment a single-arm study. This method can be utilized for medical drug/device development.

Title: Context matters: Exorcising the Ghost of Pearson from Real-World Data
Speaker: Andy Wilson (PAREXEL)

In the beginning of the 20th century, science underwent a strong anti-causal movement. It was proposed that science should be all about correlations, and causality should be radically eliminated from science and relegated to obsolescence. As Karl Pearson said, “The ultimate scientific statement of the description of the relation between two things can always be put on a contingency table.” In other words, data is all there is to science. Throughout most of the 20th century, statistical practices have embodied this associative approach to data and have mostly eschewed attribution of cause. However, even Pearson himself wrote several papers on spurious correlation – a concept that doesn’t make sense without reference to cause.

In the later part of the 20th and early part of the 21st century, we are entering what the likes of Judea Pearl are calling A Causal Revolution. There have been enormous advances in economics, epidemiology, and other fields using causal structural frameworks to estimate the impact of changing an input on an output. This ‘counterfactual’ thinking has allowed us to estimate what would have happened to an outcome ‘had we’ intervened.

One key feature that makes this possible is the (full) integration of context into the data generation model. Causal questions can never be answered by data alone. They require us to formulate a model of the process that generates the data. This is particularly important in the context of real-world data, or data generated for other purposes.

In this presentation, we will explore examples of real-world data problems, the use of causal tools such as directed acyclic graphs and marginal structural models, and the frontiers of targeted learning to bridge the gap between the causal revolution and advances in (primarily associative) machine learning systems.

In the beginning of the 20th century, science underwent a strong anti-causal movement. It was proposed that science should be all about correlations, and causality should be radically eliminated from science and relegated to obsolescence. As Karl Pearson said, “The ultimate scientific statement of the description of the relation between two things can always be put on a contingency table.” In other words, data is all there is to science. Throughout most of the 20th century, statistical practices have embodied this associative approach to data and have mostly eschewed attribution of cause. However, even Pearson himself wrote several papers on spurious correlation – a concept that doesn’t make sense without reference to cause. In the later part of the 20th and early part of the 21st century, we are entering what the likes of Judea Pearl are calling A Causal Revolution. There have been enormous advances in economics, epidemiology, and other fields using causal structural frameworks to estimate the impact of changing an input on an output. This ‘counterfactual’ thinking has allowed us to estimate what would have happened to an outcome ‘had we’ intervened. One key feature that makes this possible is the (full) integration of context into the data generation model. Causal questions can never be answered by data alone. They require us to formulate a model of the process that generates the data. This is particularly important in the context of real-world data, or data generated for other purposes. In this presentation, we will explore examples of real-world data problems, the use of causal tools such as directed acyclic graphs and marginal structural models, and the frontiers of targeted learning to bridge the gap between the causal revolution and advances in (primarily associative) machine learning systems.