This is our old blog. It hasn't been active since 2011. Please see the link above for our current blog or click the logo above to see all of the great data and content on this site.

Layups: Multilevel Modeling, NBA Style

Posted by Neil Paine on July 10, 2009

I just realized I've been derelict in my linking duties recently, because I haven't thrown any love to the Basketball Geek, Mr. Ryan Parker, for some of his posts this week on multilevel modeling. Basically, MLM is a type of regression technique that you'd use in real-world situations where contextual effects occur on several levels (hence the name) and make it difficult to assume that the errors for each coefficient are uncorrelated. And basketball, as we know too well, is a game where performance is often heavily context-driven, so MLM is certainly a method that deserves more investigation as APBRmetrics becomes more and more sophisticated. This past week, Ryan used this type of random-effects model to predict 3-point shooting ability and offensive rebounding ability based on age and past performance. Essentially it's a really fancy way of regressing to the mean, but this method also has the potential to do a lot more than that because you can theoretically control for some of those pesky contextual effects that we analysts often run into when trying to unravel a game as complex as basketball.

2 Responses to “Layups: Multilevel Modeling, NBA Style”

  1. Jump Says:

    This is my third artice on this site and I am yet to find a debate? I love a good basketball debate but I can't seem to find one at the mo??? making predictions on age and past ability is interesting, how effective is a complete other story. Going to head over to Ryan's post for a look now, hopefully I can get my basketball debate fix over there!

    Ben from the jump higher in basketball hub.

  2. Jim_Thorpe Says:

    Random-effect models are fine to get slopes but I wouldn't uses them with the objective of forecasting particular players. The fact that each player has its own effect is useful to "correct" for it and avoid biasing the "average" estimates. For instance, every player that has some ability at shooting the threes but is not the prime perimeter player of the team (Ariza, Battier, Kapono...) gets a lot of open corner threes. If you were to estimate shooting ability by percentage at that spot, you would be understimating the importance of shooting corner threes in being a good perimeter player. Since corner threes are comparable across time and leagues, it is useful to compare players and its "right" effect must be assessed. I think that most (if not all) the studies presented at this web page should use either fixed or random panel data models for identification. In this sense, I view this as a step ahead.