Graphical Causal Models with Discretized Data and Background Information

In several application areas, discretized variables represent an underlying continuous variable. For example, the level of certain medical measures can be ‘low’, ‘medium’ or ‘high’, while the underlying measure is a continuous variable. The estimation of graphical causal models for data with discretized variables leads to biased estimates and underestimated causal relations. In this work, we study the effect of incorporating background information on causal relations when estimating causal models with discretized variables. We show that incorporating background information on the relations between variables improves graphical causal model estimates in case of discretized variables. We find particularly large gains in reducing omitted causal relations and in estimating causal relations correctly. We relate these improvements to the hyperparameter choice in graphical causal models and properties of the variables in the model.