Learning causal structures with background knowledge in health data
Yiheng Liang, Armin Robert Mikler · 2015
Causal analysis plays an important role in health informatics research. Many publicly available data in health, medicine, and epidemiology are non-temporal and observational. Therefore, there is a particular field of interests to explore causal relationships in observational data. The research advance in computer science, particularly in learning Bayesian networks enables investigators to discover possible causal structures through statistical testing. However, exact learning is infeasible when the number of variables are large. On one hand, causal relationships for most variables in the dataset are unknown. On the other hand, there are a few causal relationships verified by medical experiments, confirmed from existing knowledge, or practically utilized under reasonable assumptions. The difficulty in structural learning process can be reduced when variables' ordering is known. This research presents novel methods for learning causal structures with partially known orderings of variables in combination with epidemiological background and health science research.