PhD Proposal: Towards Practical Causal Inference with Graphical Causal Models: Theory and Tools for Researchers

Other than the randomized controlled trial, causal inference does not have a rigorous theory and associated practice in common use by scientific researchers. As a result, causality is often discussed informally with reported statistics suggesting causal effect but remaining inconclusive. I propose that the theory of graphical causal models is a powerful and general approach to causal inference that can support the development of expressive, robust tools and software to make rigorous causal inference practical for non-experts in the theory. In preliminary work, I introduce a 'causal coefficient' analogue to the correlation coefficient and systematically analyze the possible relationships between correlation and causation. From this, I develop formal arguments that show while common intuition about correlation and causation is often correct, trying to reason informally about causation can still lead to wrong conclusions that require a more rigorous approach to correct. I also analyze the computational power of dynamic Bayeisan networks and conclude that the key restriction on graphical causal models - no cycles - is necessary in the sense that any cyclic extension will either have to restrict the types of random variables permitted or be undecidable. The proposed research plan includes using the theory of graphical models to develop `cookbook' approaches to reporting interval estimates for causal effect, given a formula to estimate causal effect from experimental data, and implementing an embedded language to derive such formulas. Examining Committee: Chair: Dr. James Reggia Dept rep: Dr. Mihai Pop Member: Dr. Rance Cleaveland