Falsifying causal models via nonparametric conditional independence testing

Lucas Kook, WU Vienna

Abstract: Estimating causal effects from observational data requires (i) assumptions on the underlying data-generating process, such as a graphical causal model, and (ii) an identification strategy for the causal effect of interest, such as covariate adjustment. Both (i) and (ii) typically involve untestable assumptions, making it crucial to be able to criticize or falsify the resulting effect estimates. This work proposes one way to do so: Given a putative causal model and an observational dataset, we first extract testable conditional independence relations from the causal model. We then nonparametrically test those relations, potentially falsifying the causal model, while controlling Type~I error. We illustrate the approach based on covariance measure tests, a family of regression-based nonparametric conditional independence tests, by falsifying two causal models of protein interactions using publicly available single cell flow cytometry data.