I don't have any actual real-world numbers at hand. What I'm writing is a result of 17-year experience as RF engineer at a MNO, most of time dealing with single-frequency networks (UMTS, LTE ... indeed that's mostly FDD but the interference problem is the same regardless) where interference (cosited, non-cosited, even non-cochannel) is a bad thing. And believe me, we've done our share of research into it and develop some improvements - in few rare cases we also applied custom antenna shielding. Our actions were "complaint-driven", so we never did a full "scientific" report, while we did sometimes rethink the solution if it turned out to degrade significantly the service for users who were not suffering in the first place.
But then I'll quote myself: "If original antenna is at least half-decent, then additional shielding won't change radiation pattern in main lobe direction much". The problem is, that many antennae with price tag less than outrageous, are less than half-decent (and that, by the look of it, includes the antennae shown in the presentation near time-mark mentioned). When "improving" such antennae, it surely matters quite much how you do the shielding ... I agree if done poorly, it can degrade antenna performance considerably.
The basic idea when we were designing out network was: construct 3-sector site using antennae with 65° horizontal beam width (even though the sectors were spaced at 120°). Those users in sector main direction will have really great signal (with low interference) while users in between sectors will have interfered signal either way (and performance for those users will likely be limited by interference, not by somehow lover signal level). Rationale is that what's good for sector throughput (have a part of users in great signal, the rest in a mess) is actually good for all users (those in great signal will get their service while consuming low air-time making more air-time for those users with poor signal ... if all users were in mediocre signal, sector-wide throughput might be lower).