Combining individual patient data from randomized and non-randomized studies to predict real-world effectiveness of interventions