Causal inference: Principles unifying experimental and observational accounts

Abstract

Causal reasoning is closely related to interventions. We believe that if A causes B, then a change in A leads to a change in the probability distribution of B. This interventionist view of causality is natural to many, in particular to social scientists, because it is close to the standard paradigm of experimentation. This has led some to believe that the only meaningful way to obtain any causal claim is by experimentation. Here we suggest that causal claims can also be obtained from observational data, although with more limitations than with experimental data. We claim this by showing that the interventionist view of causality is on an equal footing with searching for causal claims in observational data (the causal evidence view of causality) in terms of the criteria to assess whether there is a causal relation. The limitations of obtaining causal claims from observational data result in more uncertainty, especially because of possible confounders; this uncertainty is made explicit in results from such causal discovery. We illustrate the possibilities and limitations with observational data in an empirical example on eating disorder symptomatology.