In many areas of science, understanding dependencies in data and identifying cause-and-effect relationships is essential. Linear causal models are a powerful approach for this. In these models, random variables are expressed as linear functions of parent variables and error terms. In most cases, normally distributed error terms are considered. In this paper, however, we consider models whose error terms are not normally distributed and which are distorted by the influence of general confounding variables. The structural equations are represented using acyclic-directed mixed graphs (ADMGs). This representation provides a graphical criterion for determining the generic identifiability of such models. In practice, this criterion can be implemented in different ways, such as replacing the dependency measure with appropriate, consistent estimates. We implement this criterion in Python and form an optimization problem whose solution reconstructs the causal parameters in the best possible way under the given conditions. The performance of this estimation method is evaluated by simulations, demonstrating its robustness and applicability.
«
In many areas of science, understanding dependencies in data and identifying cause-and-effect relationships is essential. Linear causal models are a powerful approach for this. In these models, random variables are expressed as linear functions of parent variables and error terms. In most cases, normally distributed error terms are considered. In this paper, however, we consider models whose error terms are not normally distributed and which are distorted by the influence of general confounding...
»