Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation