Observability and its impact on differential bias for clinical prediction models