On Causal Inference for Ordinal Outcomes