On the Use of Generative Models in Observational Causal Analysis
This work highlights a fundamental limitation in causal inference methods, which is important for researchers and practitioners in fields like statistics and machine learning, but it is incremental as it builds on existing critiques.
The paper critiques the use of generative models in observational causal analysis, arguing that they are insufficient for inferring causal relationships because they only describe a single observable distribution and cannot account for intervention effects.
The use of a hypothetical generative model was been suggested for causal analysis of observational data. The very assumption of a particular model is a commitment to a certain set of variables and therefore to a certain set of possible causes. Estimating the joint probability distribution of can be useful for predicting values of variables in view of the observed values of others, but it is not sufficient for inferring causal relationships. The model describes a single observable distribution and cannot a chain of effects of intervention that deviate from the observed distribution.