Abstract

Unmeasured confounding is typically viewed as a fundamental obstacle to causal identification. At the same time, modern datasets are increasingly multidimensional, featuring large collections of exposures and multiple outcomes measured jointly and over time. In this talk, I argue that such multidimensional structure can substantially reshape the identification problem. Structural sparsity among many exposures can help distinguish causal effects from spurious associations. Likewise, symmetry and redundancy across multiple or longitudinal outcomes can introduce constraints that facilitate identification despite confounding. From this perspective, identifiability is not only a property of isolated treatment-outcome pairs, but emerges from the broader structural geometry of the data generating process. Multidimensional designs thus provide new avenues for discovering causal structure and restoring identification in settings where classical single treatment, single outcome paradigms would fail.