#rapm

Best way I can try to explain it to have meaning, goes back before it goes forward, so bear with me. Let's go through the evolution of the +/- stats, a bit, as it will help the explanation make more sense. 1) Raw +/-. This was the first usage. Simple concept, like hockey: add up all the points a team scores and gives up while player is on court, use that as a measure. Gives some general info, but caliber of teammates on floor with player dominates the results 2) On/off +/-. This was an attempt to separate a player a bit more from his teammates. Measure +/- when player on the court, +/- when player off court, and subtract them. Still a very simple measurement...sometimes I mentally keep track of it during a game, much like you might keep track of a player's points or rebounds. It does help a bit with isolating player from teammates, but still get lots of colinearity possible within a unit. Also (on message), susceptible to what you describe...having a better/worse cast on bench would affect the "off" part of the measure. 3) APM, RAPM, etc. This one groups all the regressions together for brevity. In this approach, you no longer start with a player's on-court and/or off-court and subtract them. Instead, you use 5-man line-up data. You take every combination of 5 players that played in a given season; also take every combination of 5 players that a unit played against; and use those as inputs into your algorithm. Details vary based on which method you're using and how you implement it, but the punch line is that the regression results correlate a player's presence on the court with a certain amount of change in his unit's interactions with opponents. And this correlation is done for everyone, taking into account the strengths of the different 5-man units AND the strengths of the different 5-man opponents. APM/RAPM come with their own sets of warts that need to be understood, but the point here is that the method is supposed to account for strength of teammates and opponents, such that the quality of your cast shouldn't materially affect the bottom line. Again, and I stress this every time we get into stat discussions, it isn't perfect...it can be very noisy, for example, and requires lots of data to increase accuracy. But, this approach at least minimizes the type of teammate dependence that you bring up better than any other impact approach that we currently have access to in the public.