Assuming even distribution of four hex values: 16^4 = 65k potential IP collisions. From my quick skim of the paper, the authors made some assumptions about posting tendencies (frequent posters likely to comment on multiple topics) and looked for enriched patterns of IP addresses. An IP address assigned to multiple topics within a short timeframe is more likely to be real. As a control set, they took a different four values of the hash function (eg true function samples [10:14], false set took [11:15]) and used that as their statistical threshold.
I think there's where the anonymity claims might have come from