This thesis investigates the central notion of model equivalence in causal discovery. If two models are equivalent, then no method based solely on observational data can, or should, distinguish between them. Consequently, characterizing model equivalence is fundamental as principled discovery methods rely on structural identifiability results. Specifically, this thesis considers linear non-Gaussian causal models with general latent confounding, where latent effects are allowed to be nonlinear. In this important yet largely unexplored setting, model equivalence remains poorly understood, and no graphical characterization is currently known. This longstanding challenge has been a major obstacle to progress in causal discovery under general latent confounding. We address this gap by establishing, to the best of our knowledge, the first graphical characterization of model equivalence in this setting and by developing algorithms to decide equivalence in challenging cases.     «

This thesis investigates the central notion of model equivalence in causal discovery. If two models are equivalent, then no method based solely on observational data can, or should, distinguish between them. Consequently, characterizing model equivalence is fundamental as principled discovery methods rely on structural identifiability results. Specifically, this thesis considers linear non-Gaussian causal models with general latent confounding, where latent effects are allowed to be nonlinear. I...     »