Advancing Causal Decomposition and Discovery for Real-World Applications

How can researchers reliably apply causal methodology in practice? This talk considers two instances -- causal decomposition and causal discovery -- and develops statistical methods to support their use in practice.

In the first part of my talk, I draw on the causal decomposition framework introduced by Vanderweele and Robinson (2014) to study racial disparities in the grant peer review process at the NIH. The key research questions are: 1) Which hypothetical interventions on the attributes of principal investigators such as their career stage or institution prestige could eliminate or reduce racial disparities in selection into panel discussion? 2) What are the potential impacts of procedural modifications such as those proposed in the NIH's new simplified peer review framework? I apply the causal decomposition framework for the peer review setting by formalizing the assumptions and identification forms for interventions in this setting. I then extend this framework by 1) defining and identifying a new estimand of interest, the Thresholded Disparity, and 2) developing Bayesian estimation methods. Our findings indicate that a major simplification in NIH's new peer review framework, intended to address reputation bias, is unlikely to reduce racial disparities in selection into discussion.

Next, I turn to causal discovery methodology, which aims to infer causal relationships between variables using observational data. Functional causal discovery methods, such as those based on the Linear Non-Gaussian Acyclic Model (LiNGAM) Shimizu et al. (2006), rely on structural and distributional assumptions to determine causal directionality. However, in practice, these methods typically do not provide statistical guarantees for directionality estimates. In addition, little is known about the impact of assumption violations. In my second project, I formalize a simple hypothesis testing-based approach for bivariate causal discovery that repurposes a goodness-of-fit and independence test. I highlight the statistical guarantees this approach provides. Building on this, I introduce the conceptual framework for the third project that explores a notion of statistical power in bivariate functional causal discovery, particularly when assumptions are violated. This framework uses subsampling to evaluate how assumption violations affect causal directionality estimates. I conclude by outlining expected contributions and future work.