Improving inference from EHR data: application of causal inference to pancreatic ductal adenocarcinoma

Electronic health records (EHRs) provide a voluminous amount of data for research. However, potential biases threaten the internal and external validity of observational studies re-using EHR data for research. We visualized potential biases introduced into the EHR during its data generation process and employed directed acyclic graph (DAG), a causal inference tool, to minimize potential confounding and selection biases. We developed a framework to generate causal DAG and applied it to analyze pancreatic ductal adenocarcinoma.