Below is a snapshot of the Web page as it appeared on 4/6/2011 (the last time our crawler visited it). This is the version of the page that was used for ranking your search results. The page may have changed since we last cached it. To see what might have changed (without the highlights), go to the current page.
Bing is not responsible for the content of this page.
APBRmetrics :: View topic - Predictors of Adjusted Offensive and Defensive +/- Ratings
APBRmetrics Forum Index APBRmetrics
The statistical revolution will not be televised.
 
 FAQFAQ   SearchSearch   MemberlistMemberlist   UsergroupsUsergroups   RegisterRegister 
 ProfileProfile   Log in to check your private messagesLog in to check your private messages   Log inLog in 

Predictors of Adjusted Offensive and Defensive +/- Ratings
Goto page 1, 2, 3  Next
 
Post new topic   Reply to topic    APBRmetrics Forum Index -> General discussion
View previous topic :: View next topic  
Author Message
Dan Rosenbaum



Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina

PostPosted: Wed Aug 10, 2005 2:10 pm    Post subject: Predictors of Adjusted Offensive and Defensive +/- Ratings Reply with quote

(This was inspired by the thread started by Nikos, but once this post got really long, I figured it probably made sense to start a new thread.)

Below I have included regression results relating my adjusted offensive and defensive plus/minus rating to various box score statistics. These are not quite the series of regressions I used to produce my statistical plus/minus ratings, but I think these should be useful in talking about how box score statistics related these offensive and defensive adjusted plus/minus ratings.

Here are some important points. The way you should read the coefficients is the following. In the offenive regression, the offensive adjusted plus/minus rating increases by 0.70 points per 40 minutes with every one point increase in points per 40 minutes, holding all of the other variables (including true shot attempts) constant.

1. The offensive regression has a much higher R-squared than the defensive regression, which confirms the expectation that our box score statistics do a better job explaining offensive effectiveness.

2. There is slightly more variation in offensive ratings than defensive ratings.

3. The results suggest that holding the other variables constant, a player with better than a 37% true shooting percentage tends to increase offensive efficiency. My interpretation is that the reason this is not higher is that players who can create shots are valuable, i.e. that most players will see their true shooting percentage fall as their true shot attempts increase.

4. Interestingly, even after accounting for points scored, players with more three point attempts tend to have higher offensive adjusted plus/minus ratings. This suggests that in addition to the points they score, the ability of three point shooters to spread the floor is very important for offenses. Also, note that more three point attempts is not associated with worse defense. It does not appear that the long rebounds from missed three pointers is typically leading to easy transition points.

5. Players who go to the line more, holding the other variables constant, tend to be more effective on offense and defense. In fact, the effect is larger on defense.

6. As expected, offensive rebounds predict offensive effectiveness and defensive rebounds predict defensive effectiveness. Note, however, that the offensive rebounds appear to be more important.

7. Holding the other variables constant, players who turn the ball over tend to be not only less effective offensive players, but also less effective defenders.

8. Holding the other variables constant, steals are almost as important a predictor of offensive effectiveness as they are of defensive effectiveness. Part of this may be that steals often generate high percentage scoring opportunities for teammates, but I wonder if part of this isn't that players who get steals tend to do a better job of spacing and collecting loose balls on the offensive end. This may help explain why steals may not be a good predictor of team success, but are important at the individual level. Players who steal the ball a lot may do a better job helping their teammates avoid turnovers and bad shots.

9. For the whole sample there is not a strong relationship between assists and defensive effectiveness. But when I limit the sample to big men, I find that better assisters tend to better defensive players, holding the other variables constant.

10. Holding the other variables constant, blocks predict defensive effectivenss, but not offensive effectiveness.

11. Holding the other variables constant, players with more personal fouls tend to be more effective defenders, with little effect on offensive effectiveness.

12. Holding the other variables constant, players who play more minutes per game tend to be more effective on offense and defense. This is pretty strong evidence that coaches do observe contributions that players make that are not picked up in box score statistics.

Code:
                                           The SAS System       14:08 Wednesday, August 10, 2005  63

Model: MODEL1 
Dependent Variable: OFF1                                               

                                        Analysis of Variance

                                           Sum of         Mean
                  Source          DF      Squares       Square      F Value       Prob>F

                  Model           12 9291008.6501 774250.72084      118.118       0.0001
                  Error         1081 7085843.2932 6554.8966634
                  C Total       1093 16376851.943

                      Root MSE      80.96232     R-square       0.5673
                      Dep Mean      -0.42353     Adj R-sq       0.5625
                      C.V.      -19115.86189

                                        Parameter Estimates

                                 Parameter      Standard    T for H0:               
                Variable  DF      Estimate         Error   Parameter=0    Prob > |T|

                INTERCEP   1     -7.056284    0.61411305       -11.490        0.0001
                PTS        1      0.702730    0.06387650        11.001        0.0001
                TSA        1     -0.525243    0.06276998        -8.368        0.0001
                FTA        1      0.083834    0.06323568         1.326        0.1852
                TA         1      0.327152    0.04249266         7.699        0.0001
                AS         1      0.640857    0.04863086        13.178        0.0001
                OR         1      0.733202    0.10084425         7.271        0.0001
                DR         1     -0.138614    0.05560930        -2.493        0.0128
                TO         1     -1.042591    0.14327755        -7.277        0.0001
                ST         1      0.713849    0.14956205         4.773        0.0001
                BK         1     -0.111075    0.10316250        -1.077        0.2819
                PF         1     -0.093128    0.08545434        -1.090        0.2760
                MPG        1      0.043603    0.01161761         3.753        0.0002

                                           The SAS System       14:08 Wednesday, August 10, 2005  64

Model: MODEL2 
Dependent Variable: DEF1                                               

                                        Analysis of Variance

                                           Sum of         Mean
                  Source          DF      Squares       Square      F Value       Prob>F

                  Model           12 4432015.5876  369334.6323       48.579       0.0001
                  Error         1081 8218599.2007 7602.7744687
                  C Total       1093 12650614.788

                      Root MSE      87.19389     R-square       0.3503
                      Dep Mean       0.30028     Adj R-sq       0.3431
                      C.V.       29037.83409

                                        Parameter Estimates

                                 Parameter      Standard    T for H0:               
                Variable  DF      Estimate         Error   Parameter=0    Prob > |T|

                INTERCEP   1     -3.683703    0.66138060        -5.570        0.0001
                PTS        1     -0.067574    0.06879300        -0.982        0.3262
                TSA        1     -0.105195    0.06760132        -1.556        0.1200
                FTA        1      0.179179    0.06810286         2.631        0.0086
                TA         1      0.007954    0.04576327         0.174        0.8620
                AS         1      0.035210    0.05237392         0.672        0.5015
                OR         1     -0.126936    0.10860612        -1.169        0.2428
                DR         1      0.393748    0.05988948         6.575        0.0001
                TO         1     -0.382290    0.15430545        -2.477        0.0134
                ST         1      1.080512    0.16107366         6.708        0.0001
                BK         1      1.014717    0.11110280         9.133        0.0001
                PF         1      0.309126    0.09203166         3.359        0.0008
                MPG        1      0.057194    0.01251181         4.571        0.0001

Dependent variable first regression - adjusted offensive plus/minus rating
Dependent variable second regression - adjusted defensive plus/minus rating
PTS - points per 40 minutes
TSA - true shooting attempts per 40 minutes
FTA - free throw attempts per 40 minutes
TA - three point attempts per 40 minutes
AS - assists per 40 minutes
OR - offensive rebounds per 40 minutes
DR - defensive rebounds per 40 minutes
ST - steals per 40 minutes
BK - blocks per 40 minutes
PF - personal fouls per 40 minutes
MPG - minutes per game

All variables are pace adjusted and the regressions are using data from 2002-03 through 2004-05 and are weighted by minutes played.
Back to top
View user's profile Send private message Send e-mail Visit poster's website Yahoo Messenger
kjb



Joined: 03 Jan 2005
Posts: 865
Location: Washington, DC

PostPosted: Wed Aug 10, 2005 3:01 pm    Post subject: Reply with quote

I think this is GREAT stuff. I'll let others comment on the math, but I love stuff like:

Quote:
7. Holding the other variables constant, players who turn the ball over tend to be not only less effective offensive players, but also less effective defenders.
Back to top
View user's profile Send private message AIM Address Yahoo Messenger
Eli W



Joined: 01 Feb 2005
Posts: 402

PostPosted: Wed Aug 10, 2005 3:06 pm    Post subject: Reply with quote

Sorry if this is obvious, but what's the formula for true shooting attempts? I haven't seen that terminology.
Back to top
View user's profile Send private message Send e-mail Visit poster's website
Dan Rosenbaum



Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina

PostPosted: Wed Aug 10, 2005 3:12 pm    Post subject: Reply with quote

John Quincy wrote:
Sorry if this is obvious, but what's the formula for true shooting attempts? I haven't seen that terminology.

Good question. It is just the denominator of true shooting percentage.

True shooting attempts = FGA + 0.44*FTA
True shooting percentage = (PTS/2)/(FGA + 0.44*FTA)

Sometimes I think I also refer to "true shooting attempts" as "true shot attempts."
Back to top
View user's profile Send private message Send e-mail Visit poster's website Yahoo Messenger
olcoach43



Joined: 10 Aug 2005
Posts: 28
Location: Indianapolis, Indiana

PostPosted: Wed Aug 10, 2005 3:45 pm    Post subject: Defensive predictor #7 Reply with quote

Dan I assume that these ratings are a function of what happens to the team when a given player is on floor? (If not can you clarify?)
If so, the defensive performance is most likely negative when a poor ballhandler (turnovers) is on the floor, because of two perfectly logical factors:
1. There may well be a points off of turnovers factor, depending upon where and what type of turnover occurs. Eg, a quick basket
2. Even if the turnovers do not lead directly to quick baskets, points per possession will suffer while turnovers are occuring, and chemistry and defensive effort will break down within the team concept.

On the bench we would be focused upon breaking a negative momentum and assistants would be hollering for me to get his a** out of there! Smile
Back to top
View user's profile Send private message Send e-mail Visit poster's website
Eli W



Joined: 01 Feb 2005
Posts: 402

PostPosted: Wed Aug 10, 2005 4:03 pm    Post subject: Reply with quote

Dan, this stuff is great. I've already plugged in the coefficients to an Excel formula and gone to work. One initial thing I noticed was how poorly Ben Gordon came out defensively. I didn't use any pace adjustments, but I have him ranking third worst in the league for 04-05 (behind Troy Hudson and DeShawn Stevenson). You have Hudson and Stevenson both ranked in the 0 percentile in defensive statistical plus/minus in your 82games article, but Gordon is listed as in the 52nd percentile. Is this big difference for Gordon due to the fact that you used slightly different regressions (or due to pace adjustments)? Or did I err in my calculations?
Back to top
View user's profile Send private message Send e-mail Visit poster's website
olcoach43



Joined: 10 Aug 2005
Posts: 28
Location: Indianapolis, Indiana

PostPosted: Wed Aug 10, 2005 4:17 pm    Post subject: Ben Gordon Reply with quote

John,
I am not capable of doing the statistical work you just described. However, I am heartened by your findings. My work shows Gordon to be one of the worst "off the ball" players in the NBA, ranked 5th from the bottom of all 500+ players.
His A/TO ratio is .88 and he contributes virtually nothing in any other categories. He is a known poor defender, so I have been
wondering about this issue as well.


Last edited by olcoach43 on Wed Aug 10, 2005 4:42 pm; edited 1 time in total
Back to top
View user's profile Send private message Send e-mail Visit poster's website
jkubatko



Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH

PostPosted: Wed Aug 10, 2005 4:31 pm    Post subject: Reply with quote

Dan, did you look into the problem of multicollinearity at all? If you are just interested in obtaining predictions, then multicollinearity is not a big deal. But if your goal is to understand how each predictor influences the response, then multicollinearity is a *big* problem. Based on your initial post it seems like you are interested in the latter.
_________________
Regards,
Justin Kubatko
Basketball-Reference.com
Back to top
View user's profile Send private message Send e-mail Visit poster's website
Dan Rosenbaum



Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina

PostPosted: Wed Aug 10, 2005 6:28 pm    Post subject: Reply with quote

John Quincy wrote:
Dan, this stuff is great. I've already plugged in the coefficients to an Excel formula and gone to work. One initial thing I noticed was how poorly Ben Gordon came out defensively. I didn't use any pace adjustments, but I have him ranking third worst in the league for 04-05 (behind Troy Hudson and DeShawn Stevenson). You have Hudson and Stevenson both ranked in the 0 percentile in defensive statistical plus/minus in your 82games article, but Gordon is listed as in the 52nd percentile. Is this big difference for Gordon due to the fact that you used slightly different regressions (or due to pace adjustments)? Or did I err in my calculations?

It won't give the same results for a few different reasons.

1. Pace adjustments - although for the Bulls I cannot imagine this matters much.

2. Different coefficients - First, I use a slightly different specification of variables, which probably makes a slight difference. Second, I use the results from several regressions to get the statistical plus/minus ratings. One regression uses the whole sample, but others limit the sample by position or by true shooting attempts. Thus, the coefficients that I use are different than those in the post above.

3. I adjust the statistical defensive plus/minus ratings by team so that they add up to the defensive rating. I do the same with the adjusted plus/minus ratings, but those adjustments are small. But the stats do such a poor job predicting defensive effectiveness, the adjustment is sometimes pretty large, especially for a team like the Bulls in 2004-05.

4. Remember my ratings are predictions for 2005-06. And remember I find that rookies improve a lot. Throwing in the improvement due to getting older as well and I bet that adjustment adds a point or two to Gordon's rating.

Another thing to remember is that the percentiles that I give are weighted and by position. Shooting guards tend to have low defensive adjusted plus/minus ratings, so overall Gordon is rated much lower than the 52 percentile with the statistical rating.

But none of this takes away that in 2004-05, Gordon's defensive adjusted plus/minus rating was phenomenal. This is not affected by these adjustments above and it was higher than what I am projecting for 2005-06. So yes, the stats are saying the Gordon is a so-so defender or worse if I don't make the adjustments above, but the Bulls played great defense when he was in the game in 2004-05.
Back to top
View user's profile Send private message Send e-mail Visit poster's website Yahoo Messenger
Dan Rosenbaum



Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina

PostPosted: Wed Aug 10, 2005 6:39 pm    Post subject: Reply with quote

jkubatko wrote:
Dan, did you look into the problem of multicollinearity at all? If you are just interested in obtaining predictions, then multicollinearity is not a big deal. But if your goal is to understand how each predictor influences the response, then multicollinearity is a *big* problem. Based on your initial post it seems like you are interested in the latter.

In general, I am mostly concerned about prediction with these results, but for the discussion in this thread I am trying to make statements about how each variable is correlated with the adjusted plus/minus ratings, holding the other variables constant.

Multicollinearity, in the classic case, does not bias coefficient estimates. It just results in larger standard errors. But most of the coefficient estimates are fairly precisely estimated, so I do not really have a standard error problem. (The standard errors are probably a bit overstated, because I have not accounted for autocorrelation in the errors of multiple observations of the same player.)

Now it is possible that in cases of severe multicollinearity where the model is misspecified that multicollinearity might bias coefficient estimates. But except for the two rebounding variables, this is probably not a big problem in these regressions. So I disagree that multicollinear is a "big" problem in these regressions.
Back to top
View user's profile Send private message Send e-mail Visit poster's website Yahoo Messenger
jkubatko



Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH

PostPosted: Wed Aug 10, 2005 9:40 pm    Post subject: Reply with quote

Dan Rosenbaum wrote:
Players who go to the line more, holding the other variables constant, tend to be more effective on offense and defense. In fact, the effect is larger on defense.


Based on your SAS output, the coefficient for free throw attempts is not statistically signifcantly different from 0. Since true shooting attempts is a linear combination that includes free throw attempts, I'm wondering if the correlation between true shooting attempts and free throw attempts (which is likely to be at least 0.85) is causing this.
_________________
Regards,
Justin Kubatko
Basketball-Reference.com
Back to top
View user's profile Send private message Send e-mail Visit poster's website
Dan Rosenbaum



Joined: 03 Jan 2005
Posts: 541
Location: Greensboro, North Carolina

PostPosted: Wed Aug 10, 2005 10:27 pm    Post subject: Reply with quote

jkubatko wrote:
Dan Rosenbaum wrote:
Players who go to the line more, holding the other variables constant, tend to be more effective on offense and defense. In fact, the effect is larger on defense.


Based on your SAS output, the coefficient for free throw attempts is not statistically signifcantly different from 0. Since true shooting attempts is a linear combination that includes free throw attempts, I'm wondering if the correlation between true shooting attempts and free throw attempts (which is likely to be at least 0.85) is causing this.

First, I should add, that looking at the FTA coefficient (or the TA coefficient), holding the other variables constant, means that we are holding true shot attempts and points constant. In other words, it is giving us the effect of subsituting free throw attempts for two or three point field goal attempts without total points scored changing.

The correlation is a bit lower than 0.85 at 0.66842 and yes, that contributes to a larger standard error. But a 95% confidence interval between -0.05 and 0.22 seems reasonably precise to me.

The most highly correlated variables are points and true shot attempts, which have a correlation over 0.96. But I am able to separately identify these two variables in the offensive rating equation, but as expected I am not able to identify them in the defensive rating equation. Partially for this reason, I use a little different specification when I compute the statistical ratings.
Back to top
View user's profile Send private message Send e-mail Visit poster's website Yahoo Messenger
jkubatko



Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH

PostPosted: Wed Aug 10, 2005 10:35 pm    Post subject: Reply with quote

Dan Rosenbaum wrote:
The correlation is a bit lower than 0.85 at 0.66842 and yes, that contributes to a larger standard error. But a 95% confidence interval between -0.05 and 0.22 seems reasonably precise to me.


Ah, I forgot you were using rates per 40 minutes rather than the raw totals. That explains my overestimate of the correlation coefficient.
_________________
Regards,
Justin Kubatko
Basketball-Reference.com
Back to top
View user's profile Send private message Send e-mail Visit poster's website
jkubatko



Joined: 05 Jan 2005
Posts: 702
Location: Columbus, OH

PostPosted: Wed Aug 10, 2005 10:49 pm    Post subject: Reply with quote

Dan, have you thought about using FTA/FGA instead of FTA? I'm just curious how that would influence the results.
_________________
Regards,
Justin Kubatko
Basketball-Reference.com
Back to top
View user's profile Send private message Send e-mail Visit poster's website
Eli W



Joined: 01 Feb 2005
Posts: 402

PostPosted: Thu Aug 11, 2005 2:07 pm    Post subject: Reply with quote

What are the coefficients for estimating overall adjusted plus/minus? Are they just the sums of the offensive and defensive coefficients?
Back to top
View user's profile Send private message Send e-mail Visit poster's website
Display posts from previous:   
Post new topic   Reply to topic    APBRmetrics Forum Index -> General discussion All times are GMT - 5 Hours
Goto page 1, 2, 3  Next
Page 1 of 3

 
Jump to:  
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum


Powered by phpBB © 2001, 2005 phpBB Group