Why prefer double robust estimators in causal inference?