Please Note:Â This seminar will be given online.
|
Department Seminar Samuel Wang Link to join seminar:Â . |
Causal discovery with non-Gaussian data
Randomized controlled trials (RCT) are the gold standard for identifying causal relationships; however, in many settings RCTs are unethical, impossible, or prohibitively expensive. Thus, the problem of causal discovery examines conditions and procedures which allow recovery of causal structure from observational data. Previous work by Shimizu et al. (2006) has shown that when the data are generated by a linear structural equation model (SEM) with non-Gaussian errors and no confounding, the causal structure can be identified from population values of the observational distribution. In this talk, we will consider the case when the system may contain latent confounding and show conditions under which the causal structure is still identifiable. Time permitting, we will also discuss some recent work on uncertainty quantification for causal discovery procedures.Â