Parsimony in Causal Inference: Epistemic Justifications and Methodological Implications

In scientific inference it is commonplace to use parsimony or simplicity as a criterion for selecting theories or models: other things being roughly equal, more parsimonious theories or models are preferred. However, there is a philosophical debate over whether this preference for parsimony can actually be justified and, if so, how. This project aims to advance the study of this topic in the context of inferring relations of cause and effect. Specifically, we will examine in depth some important minimality assumptions or constraints---which manifest preferences for parsimony---in two complementary approaches to inferring causal structures from data, and leverage our theoretical findings to better understand and improve the methods in both approaches.

The first approach is known as the Boolean or logical approach to causal inference, which targets deterministic causal relations among factors and is often featured in the toolbox for qualitative data analysis in social science. This approach was inspired by the well-known INUS account of causal regularities in the philosophical literature. The "N" in "INUS" stands for "non-redundancy", which is a minimality constraint and so the main target of our critique. We will examine several variations of this constraint, and study which conception of causation and goal of inference justifies which variation. We expect to refine existing methods and invent new methods in this approach in light of the results of the study.

The second approach that concerns us is an influential approach to inferring probabilistic causal relations among random variables, which has attracted increasingly more attention in several fields, especially artificial intelligence. In this approach a number of parsimony-related assumptions are employed in various causal inference algorithms, including the Faithfulness assumption, which has been criticized or defended by several prominent philosophers of science. We recently developed a learning-theoretic justification for (a minimality component of) the Faithfulness assumption. This novel account is not only superior to standard defenses of the assumption, but also raises a number of interesting questions that call for further investigations, including especially the prospect of extending the account to address more complex problems of causal inference, such as inference of causal structures that allow latent variables, and causal inference in nonstationary or heterogeneous settings where the data generating process changes over time or across regimes. We aim to tackle these questions in this project, and produce both epistemological insights and new methods of causal discovery.