Towards development of best practice methods of causal inferenceto assess treatment selection biomarkers from non-randomizeddata