Section on Statistics in Sports – JSM 2010

Evaluating Individual Player Contributions in Basketball

Jarad Niemi∗

Abstract
This paper introduces novel methodology for evaluating player effectiveness in basketball. We
model separately each player’s offensive and defensive contributions to their team and provide es-
timates that adjust for all players on the court. These offensive and defensive contributions are
naturally combined to provide a plus-minus statistic for each player. We build a hierarchical model
that estimates the underlying distributions for player contributions and shrinks player estimates to-
ward average with more shrinkage given to players who play less. We illustrate the methodology
using data from the 2009-2010 NBA regular season. Based on this season’s data, we find the best
player in the league to be Pau Gasol who provides his team with a 10.9 ± 4.3 point differential
compared to an average player. The best offensive player is Steve Nash who provides his team with
an additional 11.3 ± 3.0 points per game and the best defensive player is Raja Bell who reduces his
opponents scoring by 12.7 ± 7.3 points per game.

Key Words: Bayesian analysis, Gibbs sampling, hierarchical model, plus-minus statistic, offensive
rating, defensive rating

1. Introduction

In team sports around the world, statistics are maintained to measure individual player per-
formance (Kubatko et al., 2007). These statistics are meant to give coaches, analysts, and
spectators an understanding of a player’s contribution to his team. For example, in basket-
ball we have field goal percentage, offensive rebounds, and assists that indicate a player’s
offensive prowess and steals, blocks, and defense rebounds that indicate the strength of
a player defensively. Analysts quickly notice short-comings of these statistics, e.g. a
risky player might have a high number of assists per game, but an even higher number
of turnovers per game. To remedy this, more statistics are typically added such as assist-
to-turnover ratio. As more statistics are added, box scores quickly become hard to inter-
pret. For this reason, attempts have been made to take combinations of box score statistics
to provide meaningful summaries of player contribution. Examples are Dave Heeren’s
NBA efficiency rating (Heeren, 1988), Dave Berri’s player value (Berri, 1999), and John
Hollinger’s player efficiency rating (Hollinger, 2005). These methods succeed in providing
a one number summary of a player’s contribution, but, as pointed out in Winston (2009)
(and references therein), often have unfortunate implications, e.g. that the worst shooter in
the NBA should shoot more to increase their rating.
Since typical box score statistics cannot capture the inherent complexities involved
in sports, we consider the plus-minus statistic. The plus-minus statistic is simply how
many points a player’s team outscores their opponents by when that player is playing.
Better players allow their team to score more points while forcing their opponents to score
fewer and therefore their plus-minus statistic will generally be larger. The main concern
with the plus-minus statistic is that it does not account for who is on the court at a given
time, but adjusted plus-minus does. As implemented by Winston (2009), the adjusted plus-
minus statistic numerically solves simultaneously for all players’ plus-minus statistics that
minimize the squared error in observed data minus predictions. The adjusted plus-minus
∗
University of California, Santa Barbara, Statistics & Applied Probability, Santa Barbara, CA 93106-3110

4914
 Section on Statistics in Sports – JSM 2010

statistic has exactly the same interpretation of the original plus-minus statistic, but accounts
for who is on the court with whom.
In a similar fashion, Rosenbaum (2004) casts the estimation of an adjusted plus-minus
into a linear regression framework. The regression takes into account the difference in
points scored while taking into account the players contemporaneously on-court and which
side is playing at home. This analysis is restricted to players who play more than 250
minutes across two seasons where a reference player replaces those playing less than 250
minutes across the seasons. The analysis provides estimates of plus-minus for each player,
but the offensive and defensive ratings have high uncertainty.
We extend this adjusted plus-minus idea to a model based framework where play-by-
play data are analyzed for offensive play and defensive play separately and then combined
to provide an overall player contribution estimate. In addition, we include all players in
the analysis, but propose a model that shrinks player estimates toward the average with
the amount of shrinkage being inversely proportional to the amount of time played. The
paper continues as follows. Section 2 introduces the play-by-play data used for the analysis
throughout. In section 3, we introduce the model and method of parameter estimation.
Section 4 provides analyses based on the play-by-play data. Section 5 briefly compares the
method to a couple of others, suggests possibilities for extending the model, and discusses
applicability to other team sports.

2. NBA Season

The data for this analysis come from a single regular NBA season which comprises 82
games for each of the 30 teams. On average there are 2̃0 substitution times which means
there are around 20 sets of 10 (not necessarily unique) players that can score points or get
scored on. Between each substitution, we record the five players for the away team, the
five players for the home team, the amount of time those players are on the court and the
number of points that are scored by the away and home teams during that time.
These play-by-play data contain all the known 10-player combinations. These com-
binations are clear whenever the on-court players and substitutions are clearly identified.
Unfortunately, the online data published by ESPN.com and NBA.com do not state substi-
tutions that occur during quarter breaks which results in player uncertainty. In some cases,
players on court can be inferred from resulting on-court actions, e.g. shots taken and fouls
committed. When the players cannot be inferred, the observations are removed and the
resulting play-by-play data is published at http://basketballgeek.com.
These data are modified so that free throws are awarded to the set of players in the
game before the foul, that resulted in the free throws, was committed. This adjustment is
required due to a nuance in the way substitutions are handled during free throws which
would assign points to players who have just entered the game (and had no role in the foul
being committed) and also would result in scores with zero elapsed time.
Table 1 provides an example of the type of data used for this analysis. In total, we have
≈32,000 observations from the 2009-2010 NBA regular season and 443 players.

3. Individual Player Contributions

The method described here considers data on a squad, set of five players, basis. At all
times two squads are concurrently on the court: a home squad and an away squad for
the home and away teams, respectively. The goal here is to extract from these squads the
contribution for each individual player after accounting for the players on-court teammates
and opponents.

4915
 Section on Statistics in Sports – JSM 2010

Table 1: Example of three observations of the 2009-2010 NBA season providing the names
of the 5 players on the court for the home and away teams, the number of seconds the
players were on the court, and the number of points scored by both the away and home
squads.
Observation # 1 2 3
Away squad Eddie House Eddie House Eddie House
Kendrick Perkins Kevin Garnett Kevin Garnett
Marquis Daniels Marquis Daniels Paul Pierce
Rasheed Wallace Paul Pierce Rajon Rondo
Ray Allen Rasheed Wallace Ray Allen
Home squad Anthony Parker J.J. Hickson Anderson Varejao
Daniel Gibson Jamario Moon Anthony Parker
LeBron James LeBron James Daniel Gibson
Shaquille O’Neal Mo Williams Jamario Moon
Zydrunas Ilgauskas Zydrunas Ilgauskas LeBron James
Seconds 28 47 12
Away points 0 1 2
Home points 2 0 0

3.1 Data generating model

We jointly consider the points scored by the home and away squads. Let yiH (yiA ) be the
number of points scored by home (away) squad i, i ∈ {1, 2, . . . , n} where there were a
total of n observations over the course of a season. We assume
     
 P  P
ind h a a h
 Y  Y
yi = (yiH , yiA ) ∼ P o yiH Hti θp ip δp ip  P o yiA Ati θp ip δp ip  (1)
 p=1  p=1

where P o(y|λ) represents the Poisson distribution with mean λ, H (A) is the average
number of points per minute for the home (away) team, hip (aip ) is an indicator of whether
player p is on home (away) squad i, i.e. hip (aip ) = 1 if yes, and 0 if no, ti is the time in
minutes between substitutions, and P is the number of players in the NBA. This approach
combines the standard generalized linear model (McCullagh and Nelder, 1999, Ch. 6)
with a Poisson process (Lawler, 1995, Ch. 3). The relationship with the generalized linear
model is apparent if, for example, we take the logarithm of the home Poisson mean and set
θj = exp(βj ) where log(Hti ) becomes the offset.
To understand the interpretation of this model, first consider the case where θip = δip =
1 ∀ i, p, i.e. all players are equal both offensively and defensively. In this case, home (away)
squad i is expected to score Hti (Ati ) points. The value for H (A) is the empirical average
points per minute for all home (away) teams over the course of the season. Since ti is the
number of minutes of the game played by squad i, the expected points scored by the home
(away) squad is proportional to the amount of time played.
In this model, the measure of player p’s offensive (defensive) contribution is θp (δp ). If
all other parameters are constant, then the multiplicative increase in offensive (defensive)
scoring while player p is in the game compared to the average player is θp (δp ). Therefore
if θp > 1 (δp < 1) that player’s contribution to the team has a positive offensive (defensive)
effect whereas if θp < 1 (δp < 1) that player’s contribution has a negative effect.

4916
 Section on Statistics in Sports – JSM 2010

3.1.1 Hierarchical structure

We assume independent Gamma priors with common shape and rate parameters for a
player’s offensive and defensive contributions, i.e. p(θp ) = Ga(αθ , αθ ) and p(δp ) =
Ga(αδ , αδ ) where p(x) = Ga(α, β) is the gamma distribution proportional to xα−1 e−xβ .
Since E[X] = α/β, the expectation of both priors is 1. Intuitively this means the average
NBA player will have θp and δp equal to 1. In addition, this prior shrinks player contri-
bution parameters toward the average with larger values of αθ and αδ resulting in more
shrinkage.
We then build a hierarchical model by assuming the non-informative prior p(αθ , αδ ) =
p(αθ )p(αδ ) = αθ−1 αδ−1 . This hierarchical structure allows the data to determine how much
shrinkage should be applied by learning the underlying distribution of the θs and δs. For
example, if the data indicate that all θs are in a small neighborhood around one, then αθ
will be large leading to small variability in offensive player contributions and therefore a
large amount of shrinkage.

3.2 Parameter estimation

The goal of these methods is to obtain the posterior distribution of the model param-
eters conditional on the observed data, i.e. p(θ, δ, αθ , αδ |y) where θ = (θ1 , . . . , θp )0 ,
δ = (δ1 , . . . , δp )0 , and y = (y1 , y2 , . . . , yn )0 . Using Bayes’ Rule and the conditional inde-
pendence assumptions, we have
 "
P n
#
Y Y
p(θ, δ, αθ , αδ |y) ∝ p(αθ , αδ )  p(θp |αθ )p(δp |αδ ) p(yi |θ, δ) (2)
p=1 i=1

where p(yi |θ, δ) is given in equation (1). Since the posterior p(θ, δ, αθ , αδ |y) is not avail-
able in an analytically tractable form, we approximate the posterior via a Markov chain
Monte Carlo (MCMC) simulation approach, namely a Metropolis-within-Gibbs algorithm
(Gelfand and Smith, 1990). Once converged, the MCMC procedure provides draws from
the posterior distribution for all parameters.
The two general steps in the algorithm are 1) jointly sampling αθ and αδ and 2) cycling
through all players jointly sampling θp and δp for that player. Due to conditional inde-
pendencies, the first step is accomplished by using two univariate random walks to sample
from p(αθ |θ) and p(αδ |δ) where the target distributions are
 αθ −1  
P P
θp  αθP αθ −1 Γ(αθ )−P
Y X
p(αθ |θ) ∝  θp  exp −αθ
p=1 p=1

and  αδ −1  
P P
δp  αδP αδ −1 Γ(αδ )−P .
Y X
p(αδ |δ) ∝  δp  exp −αδ
p=1 p=1

The second step involves cycling through players and jointly sampling their offensive
and defensive contribution parameters. The full conditional distributions for these param-
eters for player p, i.e. p(θp , δp |y, θ−p , δ−p ) where x−p indicates the vector x with com-
ponent p removed, are independently Gamma distributed with updated hyper-parameters,
i.e. Ga(θp |αθ0 , βθ0 ) Ga(δp |αδ0 , βδ0 ). The hyper-parameters are provided in equation (3) and

4917
 Section on Statistics in Sports – JSM 2010

derived in Appendix A.
n
X
αθ0 = αθ + [hip yiH + aip yiA ]
i=1
Xn
αδ0 = αδ + [hip yiA + aip yiB ]
i=1
 
n
X Y Y (3)
βθ0 = αθ + hip Hti θkhik δkaik + aip Ati θkaik δkhik 
i=1 k6=p k6=p
 
n
X Y Y
βδ0 = αδ + hip Ati θkaik δkhik + aip Hti θkhik δkaik 
i=1 k6=p k6=p

To understand the interpretation of these updated hyper-parameters, consider αθ = αδ = 0.

Then αθ0 (αδ0 ) is the points scored by (on) player p’s team while player p is on the court and
βθ0 (βδ0 ) is the expected points scored by (on) player p’s team while player p is on the
court taking into account all the players that are currently on the court and their current
contribution parameters.
Code was written in C and called from the statistical environment R (R Development
Core Team, 2009). Convergence was assessed using the method of (Brooks and Gelman,
1997) on multiple chains initialized from dispersed starting locations. This diagnostic is
available using the function gelman.diag in the package coda in R.
To place these multiplicative parameters on a scale compatible with current plus-minus
statistics, we multiply the average NBA team points per game by the parameter minus one.
For the offensive contribution, this is equivalent to finding the expected number of extra
points a team scores if an average player was replaced with this player for an entire game.
Similarly, the defensive contribution is the expected number of extra points an opponent
would score and the plus-minus contribution is the expected difference in points between
the player’s team and her opponent’s team.

4. Analysis of the 2009-2010 NBA regular season

The 2009-2010 NBA regular season was analyzed using 10 chains with initial values
overdispersed relative to the posterior. Each chain had 1,000 burn-in and 1,000 inferen-
tial iterations and no lack of convergence was detected. All results presented combine the
inferential iterations of all 10 chains.
On average this season, the home team scored 2.11 points per minute (ppm), the away
team scored 2.05 ppm, and 99.8 points per game (48 minutes) per team. Figure 1 provides
posterior estimates for the hierarchical parameters and predictive distributions for player
offensive, defensive, and plus-minus contributions. The predictive distributions indicate
that a player contributing an extra 10 points per 48 minutes is an incredible offensive player
while a defender reducing his opponent’s scoring by 10 points in 48 minutes is incredible
defensively. These estimates appear in line with other plus-minus statistics that also have
most players in the range -10 to 10.
Table 2 provides estimates and credible intervals for the top 50 NBA players in terms
of plus-minus. The table has been sorted according to the lower 2.5%-tile of the plus-
minus credible interval. This analysis suggests that Pau Gasol was the most valuable NBA
player providing a 10.90 (6.52, 15.18) point differential after adjusting for his teammates
and opposing players. This differential occurs because Gasol provides 7.42 (4.16, 10.65)

4918
 Section on Statistics in Sports – JSM 2010

p(αθ|y) p(αδ|y)

250 350 450 300 400 500

p(PPG[θ−1]|y) p(PPG[δ−1]|y) p(PPG[θ − δ]|y)

−10 0 10 −10 0 10 −10 10

Offense Defense Plus−Minus

Figure 1: Posterior distribution (top row) for prior parameters and predictive
distributions (bottom row) for a new player’s offensive, defensive, and plus-
minus contributions to a team.

4919
 Section on Statistics in Sports – JSM 2010

Table 2: Estimates and credible intervals for plus-minus, offensive, and defensive contri-
butions for the top 50 NBA players in the 2009-2010 regular season.
Player Plus-minus Offensive Defensive
1 Pau Gasol 10.90 (6.52,15.18) 7.42 ( 4.16,10.65) -3.46 ( -6.39,-0.55)
2 Dwight Howard 9.90 (6.32,13.59) 3.38 ( 0.65, 6.19) -6.52 ( -9.11,-3.91)
3 Vince Carter 10.07 (6.12,14.09) 1.09 (-1.84, 4.00) -9.00 (-11.65,-6.24)
4 Andrei Kirilenko 9.66 (4.94,14.35) 2.85 (-0.61, 6.42) -6.84 ( -9.98,-3.49)
5 Kobe Bryant 8.78 (4.87,12.65) 4.00 ( 1.19, 6.96) -4.75 ( -7.46,-2.10)
6 Anderson Varejao 8.81 (4.64,13.10) 1.74 (-1.33, 4.93) -7.10 ( -9.97,-4.21)
7 Derek Fisher 9.16 (4.54,13.63) 6.54 ( 3.28, 9.83) -2.59 ( -5.59, 0.64)
8 Deron Williams 8.38 (4.53,12.23) 6.57 ( 3.75, 9.36) -1.80 ( -4.50, 0.85)
9 Matt Barnes 8.71 (4.53,12.91) 3.77 ( 0.74, 6.88) -4.94 ( -7.78,-1.97)
10 Paul Millsap 8.45 (4.39,12.58) 5.57 ( 2.58, 8.64) -2.87 ( -5.72,-0.06)
11 LeBron James 8.03 (4.21,11.78) 4.56 ( 1.89, 7.33) -3.45 ( -6.04,-0.83)
12 Matt Bonner 9.47 (3.92,14.83) 3.58 (-0.28, 7.56) -5.82 ( -9.48,-2.13)
13 Ron Artest 7.95 (3.91,11.91) 2.68 (-0.23, 5.68) -5.28 ( -7.96,-2.45)
14 Greg Oden 11.66 (3.67,19.75) 4.68 (-1.22,10.80) -6.96 (-12.37,-1.41)
15 Ray Allen 7.36 (3.52,11.13) 2.25 (-0.45, 5.04) -5.10 ( -7.70,-2.50)
16 Mike Bibby 7.79 (3.48,12.20) 2.44 (-0.68, 5.62) -5.38 ( -8.29,-2.35)
17 Manu Ginobili 7.46 (3.36,11.58) 3.10 ( 0.04, 6.14) -4.38 ( -7.23,-1.41)
18 Chris Andersen 8.06 (3.28,12.90) 9.85 ( 6.38,13.33) 1.81 ( -1.53, 5.25)
19 Kyle Korver 9.45 (3.21,15.56) 6.95 ( 2.46,11.62) -2.51 ( -6.71, 1.85)
20 Jameer Nelson 7.71 (3.20,12.24) 1.58 (-1.74, 4.86) -6.16 ( -9.29,-2.94)
21 Kevin Durant 6.70 (3.12,10.24) 3.87 ( 1.26, 6.49) -2.85 ( -5.29,-0.36)
22 Tim Duncan 7.03 (3.05,10.93) -0.30 (-3.05, 2.49) -7.30 (-10.01,-4.55)
23 Kevin Garnett 7.41 (3.01,11.86) 1.75 (-1.48, 5.00) -5.70 ( -8.66,-2.60)
24 Jason Williams 7.55 (2.96,12.26) 4.46 ( 1.07, 7.99) -3.11 ( -6.36, 0.10)
25 Channing Frye 7.21 (2.91,11.56) 10.23 ( 7.13,13.51) 3.06 ( 0.06, 6.08)
26 Al Horford 6.69 (2.86,10.51) 3.40 ( 0.67, 6.22) -3.30 ( -5.88,-0.58)
27 Brandon Roy 6.97 (2.86,11.19) 4.39 ( 1.33, 7.47) -2.60 ( -5.42, 0.33)
28 Josh Smith 6.76 (2.83,10.67) 3.95 ( 1.16, 6.85) -2.83 ( -5.43,-0.04)
29 Ryan Anderson 8.87 (2.77,14.89) 2.21 (-2.11, 6.58) -6.65 (-10.65,-2.41)
30 Jared Dudley 7.04 (2.52,11.53) 8.47 ( 5.24,11.81) 1.44 ( -1.63, 4.57)
31 Rashard Lewis 6.54 (2.45,10.69) 2.44 (-0.57, 5.47) -4.15 ( -6.93,-1.20)
32 Nick Collison 7.01 (2.35,11.68) -1.06 (-4.40, 2.32) -8.07 (-11.24,-4.77)
33 Richard Jefferson 5.95 (2.19, 9.86) -1.08 (-3.79, 1.77) -7.05 ( -9.64,-4.37)
34 Lamar Odom 6.07 (2.13, 9.99) 1.12 (-1.66, 4.08) -4.92 ( -7.63,-2.15)
35 Kendrick Perkins 6.41 (2.08,10.80) 1.00 (-2.06, 4.20) -5.41 ( -8.40,-2.33)
36 Rajon Rondo 5.40 (1.69, 9.05) 1.40 (-1.28, 4.05) -4.00 ( -6.54,-1.30)
37 Wesley Matthews 6.64 (1.68,11.78) 6.44 ( 2.85,10.15) -0.21 ( -3.73, 3.33)
38 Jason Kidd 5.32 (1.65, 9.09) 1.97 (-0.76, 4.73) -3.38 ( -6.04,-0.75)
39 Carlos Boozer 5.48 (1.64, 9.40) 3.63 ( 0.79, 6.48) -1.81 ( -4.49, 0.88)
40 Jamal Crawford 5.60 (1.64, 9.60) 4.17 ( 1.32, 7.06) -1.42 ( -4.23, 1.38)
41 Dirk Nowitzki 5.32 (1.62, 9.10) 3.68 ( 0.99, 6.43) -1.62 ( -4.17, 1.08)
42 Steve Nash 5.57 (1.52, 9.66) 11.27 ( 8.33,14.30) 5.68 ( 2.84, 8.58)
43 Andrew Bynum 5.73 (1.28,10.37) 1.45 (-1.87, 4.83) -4.29 ( -7.51,-1.02)
44 Shawn Marion 5.24 (1.26, 9.29) 1.60 (-1.26, 4.57) -3.63 ( -6.47,-0.79)
45 Andrew Bogut 5.62 (1.22, 9.98) 0.31 (-2.79, 3.55) -5.33 ( -8.32,-2.25)
46 Joe Johnson 4.96 (1.19, 8.83) 3.00 ( 0.25, 5.77) -1.94 ( -4.61, 0.75)
47 Rodrigue Beaubois 7.87 (1.03,14.81) 5.89 ( 0.86,11.05) -1.99 ( -6.65, 2.83)
48 Leandro Barbosa 7.85 (1.01,14.66) 12.04 ( 7.13,17.11) 4.17 ( -0.40, 8.98)
49 Nene Hilario 4.89 (0.99, 8.69) 5.37 ( 2.57, 8.18) 0.50 ( -2.15, 3.26)
50 Dwyane Wade 4.82 (0.98, 8.75) 3.77 ( 0.98, 6.57) -1.05 ( -3.77, 1.69)

4920
 Section on Statistics in Sports – JSM 2010

additional points for the Lakers and restricts his opponents to 3.46 (0.55, 6.39) fewer points
per game (ppg). Since these intervals do not contain zero, Gasol’s offensive, defensive, and
plus-minus are all significantly better than average.
The 2009-2010 league MVP was LeBron James who provided an additional 4.56 (1.89,
7.33) ppg for the Cavaliers while holding his opponents to 3.45 (0.83, 6.04) fewer ppg
resulting in an 8.03 (4.21, 11.78) ppg differential. The overlapping plus-minus intervals
for James and Gasol suggest that the two are not statistically significantly different from
each other and an estimate of the probability that Gasol’s plus-minus is greater than that
for James is 84%. So although this analysis can distinguish players from average, it cannot
distinguish among the top 10-20 players in the league. In fact, Greg Oden may be the
best plus-minus player in the league owing mainly to his defense. Unfortunately injuries
limited him to 21 games in the 2009-2010 season and therefore his contributions have large
uncertainties.
The best offensive player in the league (when sorted by lower 2.5%-tile) is Steve Nash
who provides 11.27 ( 8.33,14.30) points per game for the Suns, but also allows his oppo-
nents to score an additional 5.68 ( 2.84, 8.58) points per game. The best defensive player
is Raja Bell who limits his opponents to 14.26 (7.12, 21.18) fewer points per game, but
reduces his team’s scoring by 12.70 (5.25, 19.78) points per game. In the waning moments
of a close-fought battle, these are two players a coach would want to consider for strategic
offense-defense substitutions.
Complete results for the 2007-08 through 2009-10 regular seasons are available at http:
//www.pstat.ucsb.edu/faculty/niemi/research/basketball/player-contributions/.

5. Discussion

The approach presented in this article is a hierarchical model where player contributions
for both offense and defense affect the points per minute scored by the away and home
teams. The contribution parameters as well as their distributions are estimated using a
Markov chain Monte Carlo approach. Compared to other plus-minus statistics, the advan-
tages of this hierarchical statistical modeling approach are uncertainty quantification, pa-
rameter shrinkage to the average player, estimation of the underlying distribution of player
contributions, and separate assessment of a player’s offensive and defensive contribution to
the team.
In this article, the uncertainty of player contribution parameters as well as their prior
distribution are quantified. The parameters for the prior distribution have little uncertainty,
i.e. the difference between Ga(300, 300) and Ga(400, 400) is minor on the scale of the
multiplicative effects we are estimating. In contrast, there is considerable uncertainty in
player contribution estimates especially if interest centers on distinguishing individual play-
ers. Using one season of data, the best players can be clearly distinguished from average
players, but difficulty arises in separating the top players. A next step to decrease this
uncertainty would be to create a dynamic model for the player contribution parameters to
allow for information to be shared across and, perhaps, within seasons.
An appealing aspect of utilizing a prior distribution for the player contribution param-
eters is shrinking estimates toward average. In the full conditional distributions given in
equation (3), the prior has the effect of adding αθ and αδ points for a player’s team and
his opponent’s respectively. This has little effect on players like James who were on the
court for ≈12,000 total points, but has a large effect on players like Bell who were only
on the court for ≈100 total points. On one hand, this makes Bell’s defensive ability that
much more impressive since his defensive contribution was shrunk to the average by a large
amount. On the other hand, even Bell might have gotten lucky while he was on the court

4921
 Section on Statistics in Sports – JSM 2010

and therefore a better estimate of his ability is to shrink it toward the average.
Finally, many plus-minus statistics only provide an estimate of the point differential a
player creates when they are in the game relative to an average player. In contrast, analy-
ses using player statistics often have misleading interpretations, e.g. the worst shooter in
the league should shoot more to increase their contribution. The approach described here
takes a middle ground where players themselves are modeled rather then their statistics, but
information about their offensive and defensive contributions can be inferred.

References

Berri, D. J. (1999), “Who is ‘most valuable’? Measuring the player’s production of wins
in the National Basketball Association,” Managerial and Decision Economics, 20, 411–
427.

Brooks, S. P. and Gelman, A. (1997), “General methods for monitoring convergence of

iterative simulations,” Journal of Computational and Graphical Statistics, 7, 434–455.

Gelfand, A. E. and Smith, A. F. M. (1990), “Sampling-based approaches to calculating

marginal densities,” Journal of the American Statistical Association, 85, 398–409.

Heeren, D. (1988), The Basketball Abstract, Prentice Hall, Englewood Cliffs, NJ.

Hollinger, J. (2005), Pro Basketball Forecast, Potomac Books Inc., Dulles, VA.

Kubatko, J., Oliver, D., Pelton, K., and Rosenbaum, D. (2007), “A starting point for ana-
lyzing basketball statistics,” Journal of Quantitative Analysis in Sports, 3, 1–22.

Lawler, G. F. (1995), Introduction to Stochastic Processes, Chapman & Hall/CRC.

McCullagh, P. and Nelder, J. A. (1999), Generalized Linear Models, Chapman &

Hall/CRC, 2nd edn.

R Development Core Team (2009), R: A Language and Environment for Statistical Com-
puting, R Foundation for Statistical Computing, Vienna, Austria, ISBN 3-900051-07-0.

Rosenbaum, D. T. (2004), “Measuring how NBA players help their teams win.” http:
//www.82games.com/comm30.htm.

Winston, W. (2009), Mathletics: How Gamblers, Managers, and Sports Enthusiasts Use
Mathematics in Baseball, Basketball, and Football, Princeton University Press.

4922
 Section on Statistics in Sports – JSM 2010

A. Full conditional distributions

iid
Assume θp , δp ∼ Ga(θp ; αθ , βθ )Ga(δp , αδ , βδ ) and data observed according to equation
(1).

p(θp , δp |y, θ−p , δ−p )

" n P
! P
!#
Y Y Y
∝ P o yiH ; Hti θphik δpaik P o yiA ; Ati θpaik δphik p(θ, δ)
i=1 p=1 p=1

n P
!yiH P
!
Y Y h a
Y h a
∝ Hti θk ik δk ik exp −Hti θk ik δk ik
i=1 k=1 k=1
n P
!yiA P
!
Y Y a h
Y a h
· Ati θk ik δk ik exp −Ati θk ik δk ik
i=1 k=1 k=1

·θpαθ −1 exp(−θp βθ ) δpαδ −1 exp(−δp βδ )

 
n h yiH i n n
h a
Y X Y h a
X Y h a
∝ θp ip δp ip exp −θp hip Hti θk ik δk ik − δp aip Ati θk ik δk ik 
i=1 i=1 k6=p i=1 k6=p
 
n h yiA i n n
a h
Y X Y a h
X Y a h
· θp ip δp ip exp −θp aip Ati θk ik δk ik − δp hip Hti θk ik δk ik 
i=1 i=1 k6=p i=1 k6=p

·θpαθ −1 exp(−θp βθ ) δpαδ −1 exp(−δp βδ )

 
Pn n
hip yiH +aip yiA
X Y h a
Y a h
i=1
∝θp exp −θp hip Hti θk ik δk ik + aip Ati θk ik δk ik 
i=1 k6=p k6=p
 
Pn n
hip yiA +aip yiH
X Y a h
Y h a
i=1
·δp exp −δp hip Ati θk ik δk ik + aip Hti θk ik δk ik 
i=1 k6=p k6=p

·θpαθ −1 exp(−θp βθ ) δpαδ −1 exp(−δp βδ )

  
n
αθ + n
i=1 [hip yiH +aip yiA ]−1
P X Y Y
h a a h
∝θp exp −θp βθ + hip Hti θk ik δk ik + aip Ati θk ik δk ik 
i=1 k6=p k6=p
  
n
αδ + ni=1 [hip yiA +aip yiH ]−1
P X Y Y
a h h a
·δp exp −δp βδ + hip Ati θk ik δk ik + aip Hti θk ik δk ik 
i=1 k6=p k6=p

4923