Causal sufficiency is an assumption made in causal inference stating that we have measured all common causes of variables in the model. In this short tech report, we explore the conditions under which it’s safe to make this assumption, the challenges of causal insufficiency, and we hint at methods for managing causal insufficiency in real-world applications.
At causaLens, we construct and process large datasets in order to derive a deep, knowledge-based, understanding of the world. Unlike research involving curated synthetic datasets, many relevant features of the world can not be recorded thus leading to a fundamental challenge. Namely that causal analysis may be confounded by the presence of unobserved variables which simultaneously influence putative causal and effectual variables. This is known as the causal sufficiency problem. In this short report, we explore its technical definition and intuitive meaning, and outline techniques that can be leveraged to circumvent this critical issue in real-world modeling.
