Posted by Neil Paine on July 9, 2011
Here's a quick-n-dirty study I ran this morning... The idea is this:
Some stats seem to be more correlated with a player's role than his actual skill. Take a player out of the role, plug another similar player in, the new player produces just like the old one (and the old one can't "take the stats with him" to his new destination).
How can we quantify this, though? Well, let's identify players whose circumstances changed. I took every team since 1978 and assigned its players to 10 "roles" -- primary PG, backup PG, primary SG, etc. -- based on my detailed position file and where the players ranked on the team in terms of minutes played. I then isolated every player in that sample who:
- Played at least 500 minutes in back-to-back seasons
- Was between age 24 and 34 in back-to-back seasons (to filter out potential aging effects)
- Moved to a new "role"
- Was replacing a player who played >= 500 MP in the role and was between age 24 and 34 the previous season
This leaves us with 1,866 player-seasons to look at. For each of those, I need to know which predicts the player's performance in Year Y better -- his own stats from Year Y-1, or the Year Y-1 stats of the player whose role he took?
|Category||Player Y-1||Role Y-1|
The second column is the correlation coefficient between the player's rate in Year Y and his rate in Year Y-1. The third column is the correlation between the player's rate in Year Y and the rate the team got from his role in Year Y-1. Categories are roughly in order from least to most role-dependent.
As you can see, component stats like Usage%, Steal%, Turnover%, and Block% are fairly stable year-to-year and are explained far more by the player's previous performance than the role he plays, suggesting that those metrics capture underlying skill in addition to the player's role. At the other end of the spectrum, we see that defensive box-score estimators like Defensive Rating/Def. Win Shares per 48 minutes are heavily influenced by the team and role, which makes sense given that there is a large team component in DRtg.
Also of note: Although it's very stable year-to-year, a great deal of Assist% is role-dependent, as is also the case with the rebound percentages. This seems to confirm Phil Birnbaum's research that much of rebound rate (particularly defensive rebound%) is more a product of team roles and positional designations than actual skill. And the various offensive efficiency metrics (ORtg, Floor%) fall somewhere in between, capturing more skill than Ast% or the rebound percentages, but less than Usage%, Steal%, Turnover%, and Block%.