Better than average? When can we say that subsampling of items is better than summary statistics representations?