Explainability of Machine Learning Models under Missing Data