I actually think it has less to do with "sheltered players" and more to do with what I referenced above.
When people conduct large-scale studies of NHL players, they impart upon the study an unavoidable selection bias and survivor bias. The players have to be good enough that they can play enough NHL minutes to be included in the study in the first place. Thus, we learn certain principles about this elite group and assume that what is true in the aggregate applies to all players who step on the ice in the league, and this is a fundamental error I see made very, very often in the analytics community. You learn something interesting, and it is a genuinely interesting and useful idea that, for example, you can expect NHL players to have their PDO regress towards 100. However, this is only true for players good enough to have been included in the original data-set for whom this is true! Thus, you make a fundamental error when trying to apply this general principle to any player who steps onto the ice.
We simply cannot assume that the same basic principles we have learned studying analytics for established NHL players also apply to the Patrick Wiercoch, Derrick Pouliot, Nic Dowd, Linden Vey level of players who are, at best, AHL-NHL tweeners and at worst something much less than that. Everything we know about analytics, CF, on-ice save%, etc. needs to be very carefully applied when talking about guys who have not established themselves to be quality NHL players. If I stepped on the ice, my analytics would be unfathomly terrible, and nobody would try to dismiss my terrible PDO or on-ice save% or +/- as having anything to do with luck. It has everything to do with me not being an NHL player.
