When Real World Data are Insufficient for Causal Inference