Comments on: Calculating Adjusted Plus/Minus

By: cherokee_acb

cherokee_acb — Tue, 03 Jun 2008 17:22:03 +0000

I’ve just found that I don’t have an up-to-date script valid for the 2008 season format. If I decide to write it, I’ll share. I’ll offer another tip instead: in R, use “as.formula” and “paste” (see an example in the APBRmetrics thread linked above) to dynamically construct the regression formula from the data.

By: Eli

Eli — Mon, 02 Jun 2008 20:31:41 +0000

Another tip worth mentioning is that if when creating the CSV you use all players rather than just the qualifying (non-reference) ones, then you can later run regressions with different minutes cutoffs just by changing the regression formula you use (which is a lot quicker than having to rebuild the CSV over again).

By: Eli

Eli — Mon, 02 Jun 2008 18:42:34 +0000

Thanks Cherokee, I’ll try that.

That makes sense, Dave. I’ll update the post with that alteration. Though any centering is going to be somewhat arbitrary regardless.

By: Dave Lewin

Dave Lewin — Mon, 02 Jun 2008 18:25:48 +0000

Eli,

When adjusting the ratings so league average is zero you want to use the minute weighted average, so that the average minute played has an adjusted plus-minus of zero, not the average player.

By: Neil

Neil — Mon, 02 Jun 2008 17:33:23 +0000

Great work, Eli… I’m bookmarking this page for the day when I finally take the plunge and try it for myself. :)

By: JTapp

JTapp — Mon, 02 Jun 2008 16:28:02 +0000

Thanks for the transparency of your work here, it is really refreshing. Sagarin & Winston were charging something astronomical just to get the details of how they calculated their ratings.

By: cherokee_acb

cherokee_acb — Mon, 02 Jun 2008 15:03:31 +0000

Eli, what I’ve done is what Lior suggested, a single regression, twiche the independent variables, twice the observations. And you need to think about what the neutral ORtg is, and how to insert home advantage into the model (a dummy variable should suffice for this).

Regarding using a script to create the observations, that’s what I do, and it’s way faster than Excel (what takes the most time in my computer is the actual regression, I blame it to memory issues). At
http://sonicscentral.com/apbrmetrics/viewtopic.php?p=10823
you can see a Perl script I used for 2006 data. With little changes, it can handle the current basketballvalue format. I may post the updated script later.

By: Eli

Eli — Mon, 02 Jun 2008 12:50:56 +0000

Ah, so maybe two separate regressions, one with ORtg as the dependent variable and one with DRtg, each with twice as many independent variables? That sounds like it might work. In the Home ORtg one, the five home offensive players would get a 1, the five away defensive players would get a -1, and the five home defensive players and the five away offensive players would all get a 0. I’ll try that out later. (Or maybe you’re saying there should be one regression but with the number of observations doubled?)

By: Lior

Lior — Mon, 02 Jun 2008 11:54:24 +0000

I think the problem you’re having with separating the defensive and offensive values is that in addition to doubling the number of independent variables, you should also think of each data line as two observations: rather than using MARGIN you should write two linear relationships: one using the home team offensive ratings and the away team defensive ratings to approximate the home team efficienty during that “mini match-up”, the other replacing “home” and “away”. Given the noisiness of the usual calculation you will probably need to increase the number of reference players.

By: Ryan J. Parker

Ryan J. Parker — Mon, 02 Jun 2008 05:07:03 +0000

I never realized you could fit adjusted +/- with lm(). I thought some weird variant of glm() would be needed. This will be a handy bookmark to have.