Because this is looking at median score. Lots of people submit PG links as soon as they show up, but only one or two of those submissions will make it to the front page. If more than half of PG links have a score of 0–5, then the median will be in that range as well.
So HN itself does the merging and the raw dataset still includes the numerous duplicate submissions then? If this is the case it's not just sources with a lot of content like medium.com, github.com, nytimes.com being dragged down, it's any popular source.
