82games plusminus is weird because it subtracts on court from off court, which makes backup quality too important. For all the data 82games has, that’s really a mediocre way of doing it.

Adjusted compares a player’s plusminus to the entire league, which in a way means it compares it to nobody. It separates a player from his teammates, from his backups, from his opponents.

Imagine you want to figure out what makes a bike go fast. You try different seat heights, pedaling speeds, backwards/forwards pedaling, gears, and you jot down how fast you rode in each trial. Imagine you then put all that data into a spreadsheet, where speed is in one column, and then all the other variables are in other columns, with one row equaling one trial.

Using a regression, you can separate the impact of each variable. You’d see right away that seat height means basically nothing, pedaling backwards slows you considerably, but pedaling speed and gear can have the biggest positive effect.

APM does the same thing, but with basketball. The 10 players on the court (five per team) in any shift are your input variables, the change in the score for that shift is the output. Put every shift in the NBA into a spreadsheet and you’ve got about 12 million cells of data. Run the regression and it sorts out which players most positively and negatively affect the score.

If in an NBA season, every possible 5 on 5 was used, we’d be able to figure out how much each player’s effect is. But because players are typically trapped on a single 12 man roster, there’s trouble in separating how much positive/negative value to give to each player. That’s why you have the 5-6 noise levels in one year, and Telfair rising into the top 10.

Give it two to three years of data, and we can start figuring out who’s been the most effective. The supercomputer’s on its way.