Treatment Effect Estimation with Unobserved and Heterogeneous Confounding Variables