Estimating Heterogeneous Treatment Effects with Observational Data