RACIAL DISCRIMINATION AMONG NBA REFEREES
JOSEPH PRICE AND JUSTIN WOLFERS
The NBA provides an intriguing place to assess discrimination: referees and
players are involved in repeated interactions in a high-pressure setting, with referees
making split-second decisions that might allow implicit racial biases to become
evident. We find that more personal fouls are awarded against players when they
are officiated by an opposite-race officiating crew than when they are officiated by
an own-race refereeing crew. These biases are sufficiently large so that they affect
the outcome of an appreciable number of games. Our results do not distinguish
whether the bias stems from the actions of white or black referees.
I. INTRODUCTION
Does race influence our evaluation of others? We provide new
evidence on racial biases in evaluation by examining how the number
of fouls awarded against black and white National Basketball
Association (NBA) players varies with the racial composition of
the refereeing crew. Our setting provides intriguing insights into
own-race bias. Relative to social, judicial, or labor market settings,
the evaluators in our sample are a particularly expert group, with
substantial experience, continual feedback, and large incentives
to be accurate. NBA Commissioner Stern has claimed that these
referees “are the most ranked, rated, reviewed, statistically analyzed
and mentored group of employees of any company in any
place in the world.”
NBA referees are effectively randomly assigned to each game.
Moreover, the number of games played is large, so we can assess
both a very clear baseline rate at which individual players commit
fouls and a clear baseline for the number of fouls called by different
referees. Against these baselines, we find systematic evidence of
an own-race bias. Players earn up to 4% fewer fouls or score up
∗This paper has benefited from helpful comments from David Berri, Mark
Cuban, John Donohue, Scott Drewianka, Ronald Ehrenberg, Todd Elder, Gary
Fields, Joseph Gyourko, Kevin Hallock, Christine Jolls, Lawrence Kahn, Lawrence
Katz, Alan Krueger, Lars Lefgren, David Levine, Janice Madden, Betsey Stevenson,
Matthew White, and seminar participants at Chicago Law School, Cornell,
Dartmouth, Delaware, Missouri, the National Academy of Science, Northwestern
Law School, NYU, Stanford Law School, UCLA, and Wharton and the annual meetings
of the American Economic Association, the American Law and Economics
Association, the Conference on Empirical Legal Studies, the Society of Labor
Economists, and the NBER Summer Institute. We are also grateful to various
NBA and team officials who provided background information and feedback. We
received excellent research assistance from Bryan Elliott. The authors are grateful
to the Wharton Sports Business Initiative for research funding. jpp34@cornell.edu.
C 2010 by the President and Fellows of Harvard College and the Massachusetts Institute of
Technology.
The Quarterly Journal of Economics, November 2010
1859
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1860 QUARTERLY JOURNAL OF ECONOMICS
to 21/2% more points when they are the recipients of a positive
own-race bias, rather than a negative opposite-race effect.
We find similar results when aggregating to the team level,
with the racial composition of the refereeing crew having an appreciable
effect on the probability of a team winning. In an average
game, one team plays around 15% fewer minutes with black
players than their opponents. For this team, the chance of victory
under an all-black refereeing crew versus an all-white crew differs
by about three percentage points.
The simplest interpretation of our findings is that they reflect
own-race bias, with either black or white referees (or both) favoring
players of their own race, or disfavoring those of other races,
though we are unable to make strong statements about which
type of bias is occurring. Even so, we explore several other interpretations.
Because our unit of analysis is the refereeing crew,
we explore whether these findings can be explained by changes
in crew-level dynamics, rather than simply reflecting individual
referee biases. Alternatively, it may be that the interaction of referee
and player race is relevant, not because it affects foul-calling,
but because it affects player behavior. We also assess an omittedvariables
interpretation in which players may be disadvantaged
by opposite-race referees, but this may be the product of different
playing styles of black versus white players interacting with
different refereeing styles among black versus white referees.
Although we cannot take a strong stance on the mechanisms
involved, the accumulated evidence is most consistent with our
findings being driven by own-race bias. Comparing games with
all-black and all-white refereeing crews yields findings consistent
with the rest of the sample, suggesting that the relationship
between foul-calling and the composition of refereeing crews is
driven by individual referees favoring players of their own race.
We examine a variety of player outcomes, finding little evidence of
a rise in aggressive play that might explain the rise in the number
of fouls called against players. Our findings are also robust
to both aggregation to the team level and the inclusion of a wide
range of controls, including rich controls for playing styles and
their interaction with referee race.
Related evidence suggesting a role for own-race preferences
has been documented in a range of other contexts. Donohue and
Levitt (2001) find that an increase in the number of police of a
certain race is associated with an increase in arrests of people of
the other race. Antonovics and Knight (2009) find that police are
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1861
more likely to search the vehicle of someone of a different race.
Stauffer and Buckley (2005) find that supervisors give lower performance
ratings for workers of the opposite race. Stoll, Raphael,
and Holzer (2004) find that those firms where whites are in charge
of hiring are less likely to hire black job applicants than those
where blacks control hiring.1
The advantages of our setting lie in
the process for assigning referees to games, which takes no account
of player race, thereby ensuring that our findings are not
confounded by subjects sorting to preferred evaluators, and repeated
interactions that allow for reasonably precise inferences.
Applying Beckerian taxonomy to our findings, this own-race
preference falls under the banner of taste-based discrimination.
Within this, customer-based discrimination is unlikely, as the
own-race preference continues to exist even after we hold the stadium
(and hence customer base) constant. Additionally, employer
discrimination is inconsistent with the formal incentives for accuracy
provided by the league. This suggests a referee-specific taste
for discrimination. Although explicit animus is unlikely, Bertrand,
Chugh, and Mullainathan (2005) describe an emerging literature
on implicit discrimination that points to the role that implicit associations
(such as between blacks and violence) might play in
the types of split-second, high-pressure evaluations required of
NBA referees.2
Our findings may reflect these implicit associations
varying with the race of the evaluator.
In addition, a large literature has documented substantial evidence
of discrimination within sports (Kahn 1991). This setting
has afforded useful insights largely because measures of productivity
are easily observable. Although earlier research suggested
that black NBA players suffered substantial wage discrimination
(Kahn and Sherer 1988; Koch and Vander Hill 1988), over recent
decades, these racial gaps appear to have receded, or even disappeared
(Hamilton 1997; Bodvarsson and Brastow 1999). However,
whereas these tests for discrimination typically ask whether
wages differ for blacks and whites conditional on observable game
statistics, we demonstrate that observable game outcomes are
1. Own-race bias has also been explored in judicial sentencing, yielding mixed
results (Welch, Combs, and Gruhl 1988; Spohn 1990; Bushway and Piehl 2001;
Schanzenbach 2005). A recent study by Abrams, Bertrand, and Mullainathan
(2006) uses random assignment of judges to particular cases and finds evidence of
racial biases in terms of sentencing but not evidence of own-race bias.
2. Greenwald and Banaji (1995) provide an excellent review of implicit social
cognition. Payne, Lambert, and Jacoby (2002) note that the need to make quick
judgments increases one’s susceptibility to implicit stereotyping.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1862 QUARTERLY JOURNAL OF ECONOMICS
influenced by the racial mix of the referees. Moreover, in light
of the mismatch between the composition of the players (around
four-fifths of whom are black) and their evaluators (around twothirds
of referees are white in our sample), an own-race preference
may drive an aggregate bias against blacks (or for whites).
II. BACKGROUND: BASKETBALL, THE NATIONAL BASKETBALL
ASSOCIATION, AND REFEREES
In any season, the NBA has around sixty referees, with a
crew of three referees officiating each game. Assignments of referees
to crews are made to balance the experience of referees across
games, with groups of three referees working together for only a
couple of games before being regrouped. According to the NBA, assignments
of refereeing crews to specific (regular season) games is
“completely arbitrary” with no thought given to the characteristics
of the competing teams. Each referee works 70 to 75 games each
year, and no referee is allowed to officiate more than nine games
for any team, or referee twice in a city within a fourteen-day
period. Although these constraints mean that assignment of refereeing
crews to games is not literally random, the more relevant
claim for our approach is that assignment decisions are unrelated
to the racial characteristics of either team. For example, Table I
shows that for each year in our sample, the number of white referees
is unrelated to the number of black starters. Likewise, the
Appendix shows that none of our variables have any power in
explaining the assignment of referees of each race to particular
games within each season.
Every game has an observer who meets with the referee for
a pregame discussion, observes the game, and reviews video clips
from the game with the referees afterward. These observers report
to group supervisors, who provide additional input. The director of
officiating also provides biweekly feedback to each referee on his or
her performance. There is also an informal network of monitoring
by coaches, spectators, sports analysts, and fans.
The high level of monitoring of referees naturally leads to a
high level of accountability for their decisions on the court. The
league keeps data on questionable calls made by each referee and
uses this as an input into its internal referee evaluation system.
(Unfortunately the NBA refused to share these data with us.)
These internal ratings determine which referees will officiate the
playoffs, which provide substantial additional compensation on
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1863
TABLE I
BLACK STARTERS PER TEAM AND THE DISTRIBUTION OF REFEREEING CREWS BY RACE
Black starters per team
χ2test of
0 white 1 white 2 white 3 white independencea
Season referees referee referees referees (p-value)
1991/1992 4.33 4.33 4.27 4.28 .82
1992/1993 4.20 4.20 4.26 4.25 .03
1993/1994 4.27 4.27 4.31 4.30 .80
1994/1995 4.20 4.27 4.29 4.25 .26
1995/1996 4.35 4.26 4.29 4.23 .60
1996/1997 4.11 4.17 4.19 4.17 .97
1997/1998 4.22 4.18 4.19 4.21 .98
1998/1999 4.05 4.13 4.10 4.14 .99
1999/2000 4.26 4.25 4.14 4.25 .07
2000/2001 4.15 4.19 4.22 4.18 .99
2001/2002 4.12 4.08 4.11 4.15 .82
2002/2003 4.16 4.20 4.11 4.20 .79
2003/2004 4.03 4.05 4.03 4.04 .12
Sample size (% of 668 4,928 11,580 7,350 n = 24,526
all player–games) (2.7) (20.1) (47.2) (30.0)
Notes. Each observation is a team × game observation. Sample includes all regular season NBA games
from 1991/1992–2003/2004, excluding referee strikes.
aFinal column tests: H0: #white referees is independent of #black starters.
top of the referees’ base salary. Leading referees can earn several
hundred thousand dollars per year.
III. PLAYER-LEVEL ANALYSIS
Our data contain box score information from all regular season
NBA games played from the 1991–1992 season through to the
2003–2004 season, yielding over a quarter of a million player–
game observations. For each player–game, we observe all performance
statistics (points, blocks, steals, etc.), as well as minutes
played and the number of personal fouls committed. The box
score also lists the three referees officiating each game. Although
we cannot observe the referee who blows the whistle for each
foul, our empirical strategy involves comparing the number of
fouls each player earns based on the racial mix of the referee
crew.
We coded referees as black or nonblack based on visual inspection
of press photographs of referees, supplemented by the able
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1864 QUARTERLY JOURNAL OF ECONOMICS
assistance of a former NBA referee. Our data on player race come
from a variety of sources, including Timmerman (2000), Kahn and
Shah (2005), and our own coding from past issues of the Official
NBA Register and images on nba.com. In each case, we simply
noted whether a player or referee appeared black or not. Hispanics,
Asians, and other groups are not well represented among
either NBA players or referees, and throughout the paper we refer
to nonblacks somewhat imprecisely as “white.” We also draw
information about each player’s characteristics (height, weight,
and position) from basketballreference.com; characteristics of the
game, including the home team and attendance, from the box
score; and team characteristics, including the coach’s race, from
the NBA Register. We construct a variable for whether a team was
out of contention by calculating whether there were fewer games
left in the season than the gap between that team’s victories and
the record of the eighth best team in its conference. Some of our
player-level controls also vary by game, such as whether players
were named in the starting five, their age, their experience, and
whether they were all-stars that season. Table II provides a list
of the variables used in our analysis, as well as a comparison of
the mean values between white and black players, weighting all
player-level observations by minutes played.
These summary statistics reveal that black players play more
minutes per game than white players. Black players receive about
the same number of fouls per game (2.55 vs. 2.53) as white players,
and hence they receive fewer fouls per 48 minutes played (4.33 vs.
4.97). The differences in foul rates largely reflect the fact that
white players tend to be taller, heavier, and more likely to play
center than black players.3
However, our focus is on own-race bias, which involves assessing
how these differences vary as the racial composition
3. Note that the large unconditional black–white difference in foul rates is
explained by a few observables. First, the unconditional difference:
Fouls per 48 minsit = 4.97 − 0.64 × black playeri adj. R2
= .005 n = 266,984
(.016) (.017)
Adding covariates yields
Fouls rateit = −0.017 × black playeri + 1.47 × centeri + 0.53 × forwardi + 0.025
(.017) (.032) (.021) (.003)
× height + 0.010 × weight + 0.053 × age − 0.086 × experienceit
(.0004) (.005) (.005)
− 1.366 × starter − 0.061 adj. R2
= .097 n = 266,984
(.013) (.252)
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1865
TABLE II
SUMMARY STATISTICS (WEIGHTED BY MINUTES PLAYED)
Black players White players
Mean Mean
(SD) (SD) Difference
Raw player statistics
Minutes played 30.71 27.25 3.46∗∗∗
(9.98) (10.33)
Fouls 2.55 2.53 0.02∗∗∗
(1.51) (1.54)
Points 13.24 11.07 2.16∗∗∗
(8.37) (7.54)
Player productivity: stats × 48/minutes played
Fouls 4.33 4.97 −0.64∗∗∗
(3.20) (3.93)
Points 19.76 18.45 1.31∗∗∗
(10.05) (10.11)
Free throws made 3.86 3.52 0.34∗∗∗
(3.90) (3.99)
Free throws missed 1.33 1.11 0.22∗∗∗
(1.99) (1.99)
2 point goals made 6.59 5.96 0.62∗∗∗
(3.99) (4.02)
2 point goals missed 7.30 6.42 0.88∗∗∗
(4.24) (4.36)
3 point goals made 0.91 1.00 −0.09∗∗∗
(1.63) (1.78)
3 point goals missed 1.71 1.70 0.01
(2.36) (2.50)
Offensive rebounds 2.52 2.70 −0.18∗∗∗
(2.78) (3.09)
Defensive rebounds 5.77 6.27 −0.50∗∗∗
(4.10) (4.42)
Assists 4.57 4.22 0.35∗∗∗
(4.08) (4.30)
Steals 1.66 1.48 0.18∗∗∗
(1.88) (1.93)
Blocks 1.00 1.17 −0.18∗∗∗
(1.75) (2.06)
Turnovers 2.97 2.83 0.14∗∗∗
(2.54) (2.74)
Game information
Attendance (1,000s) 16.71 16.80 −0.09∗∗∗
(3.69) (3.62)
Out of contention 0.06 0.06 0.00
(0.24) (0.24)
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1866 QUARTERLY JOURNAL OF ECONOMICS
TABLE II
(CONTINUED)
Black players White players
Mean Mean
(SD) (SD) Difference
Black coach 0.24 0.20 0.04∗∗∗
(0.43) (0.40)
Player characteristics
Age 27.90 28.00 −0.09
(4.02) (3.87)
NBA experience (yrs) 6.19 5.78 0.41∗∗
(3.74) (3.73)
All Star this year 0.13 0.09 0.04∗∗∗
(0.34) (0.29)
Center 0.11 0.34 −0.22∗∗∗
(0.32) (0.47)
Forward 0.44 0.35 0.09∗
(0.50) (0.48)
Guard 0.45 0.31 0.13∗∗
(0.50) (0.46)
Starter 0.69 0.59 0.10∗∗∗
(0.46) (0.49)
Height (in.) 78.4 80.54 −2.13∗∗∗
(3.62) (4.14)
Weight (lb.) 211.5 223.2 −11.7∗∗∗
(26.5) (29.5)
Referees
0 white referees 0.03 0.03 −0.00
(0.16) (0.17)
1 white referee 0.20 0.21 −0.00
(0.40) (0.41)
2 white referees 0.47 0.47 0.00
(0.50) (0.50)
3 white referees 0.29 0.29 0.00
(0.46) (0.46)
# white referees 2.04 2.03 0.01
(0.78) (0.78)
Sample size Total
Players 889 301 1,190
Games 13,326 13,130 13,326
Player–games 214,291 52,693 266,984
Player–minutes 5,347,290 1,082,047 6,429,337
∗∗∗, ∗∗, and ∗ Differences statistically significant at 1%, 5%, and 10%, respectively.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1867
TABLE III
DIFFERENCES IN DIFFERENCES: FOUL RATE (=48 × FOULS/MINUTES PLAYED)
Difference: Slope:
Black White black–white (black–white)/
players players foul rate %white refs
0% white refs 4.418 5.245 −0.827
(n = 7,359) (0.043) (0.094) (0.106)
33% white ref 4.317 4.992 −0.675 0.455
(n = 54,537) (0.016) (0.035) (0.038) (0.331)
67% white refs 4.335 4.989 −0.654 0.064
(n = 126,317) (0.010) (0.023) (0.025) (0.137)
100% white refs 4.322 4.897 −0.574 0.240∗∗
(n = 78,771) (0.013) (0.029) (0.032) (0.121)
Diff-in-diff
Average slope: −0.022 −0.204∗∗∗ 0.182∗∗∗
fouls/ %white refs (0.027) (0.066) (0.066)
(p = .006)
Notes. Sample = 266,984 player–game observations, weighted by minutes played. Standard errors in
parentheses.
∗∗∗, ∗∗, and ∗ Statistically significant at 1%, 5%, and 10%.
of the refereeing crew changes. Table III shows an illustrative
differences-in-differences analysis. Reading down the columns
illustrates the two ways in which these own-race biases may
emerge: they may reflect referees favoring players of their own
race, or alternatively disfavoring those of the opposite race. The
number of fouls earned by black players is, on average, roughly
the same whether the refereeing crew is predominantly white
or black. In contrast, white players earn fewer fouls under predominantly
white refereeing crews. As such, the “difference-indifference”
suggests that a player earns 0.18 fewer fouls per
48 minutes played when facing three referees of the same race
than when facing three opposite-race referees.
This analysis reveals that the bias we document primarily
affects white players.4
This is a departure from more standard
accounts of discrimination that involve whites actively discriminating
against blacks, although our setting is unusual in that
black players are the majority group. In turn, this may reflect either
white players being favored by white referees or disfavored
4. The Online Appendix includes regressions confirming that this finding is
robust to including a broad set of control variables—although one cannot simultaneously
explore this aspect of the result, and control for referee or game fixed
effects.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1868 QUARTERLY JOURNAL OF ECONOMICS
by black referees, although our identification strategy (which relies
on random assignment of refereeing crews) does not allow us
to sort out which group of referees is responsible for this bias.
The richness of these data allows us to extend this analysis
to control for the various player, team, referee, and game-specific
characteristics that might influence the number of fouls called.
Consequently, in Table IV we report the results from estimating
Foul rateigrt = β1 %white refereesg × black playeri(1)
+ β2 %white refereesg + β3 black playeri
+ β4observable playeri, gameg, player–gameig,
team–gametg, refereer characteristics
+ player fixed effectsi + referee fixed effectsr
+ season fixed effectsg
[+ observable controlsitg × %white refereesg
+ black playeri × stadiumg effects
+ playeri effects × yearg effects
+ gameg effects + gameg effects
× teamt effects] + εigrt,
where the subscripts denote a player i playing for a team t
in a specific game g officiated by referees r. The dependent
variable is the number of fouls earned per 48 minutes, and all of
our estimates weight player–game observations by the number
of minutes played. The coefficient of interest is β1, which we
interpret as the effect of opposite-race referees on a player’s foul
rate (relative to own-race referees), or the differential impact of
the racial composition of the refereeing crew on black players
relative to white players.
In the first column of Table IV, we control for time-varying
player characteristics such as age, all-star status, and whether
the player was a starter, and team-level variables such as whether
the team is playing at home, attendance, whether they are out
of contention, and whether the coach is black. These coefficients
are reported in subsequent rows. We also control for player fixed
effects (which account for both observable differences across players,
including height, weight, and position, and unobservable
differences), as well as referee fixed effects that measure the differential
propensity of each referee to call more or less fouls. We
also control for season fixed effects to account for the fact that
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1869
TABLE IV
EFFECTS OF OPPOSITE-RACE REFEREES ON FOUL RATES
Dependent variable: foul rate (= 48 × fouls/minutes)
(mean = 4.43; SD = 3.34)
Independent variables (1) (2) (3)
Black player × 0.197∗∗ 0.203∗∗ 0.181∗∗
%white refs (0.061) (0.072) (0.080)
Control variables
Age −0.728∗∗∗ −0.729∗∗∗
(0.047) (0.049)
All-Star −0.383∗∗∗ −0.429∗∗∗
(0.026) (0.063)
Starting five −0.988∗∗ −1.004∗∗ −0.775∗∗∗
(0.016) (0.040) (0.044)
Home team −0.125∗∗∗ −0.213∗∗∗
(0.012) (0.033)
Attendance (1,000s) 0.008∗∗∗ 0.004
(0.002) (0.005)
Out of contention −0.127∗∗ −0.136∗
(0.027) (0.071)
Black coach −0.107∗∗∗ −0.080∗∗
(0.017) (0.040)
R2 .18 .18 .28
Other controls
Referee, year, and player
√ √ √
fixed effects
Player characteristics ×
√ √
%white refs
Full set of fixed effects
√
Notes. Sample = 266,984 player–game observations, weighted by minutes played (standard errors in
parentheses). Each column reports the results of a separate regression. All specifications control for the
observable variables shown (and missing coefficients reflect the fact that some controls are unidentified in the
presence of perfectly collinear fixed effects.) The second and third columns add further controls to account for
a player’s on-court role, including height, weight, position, experience, and sample averages of assists, blocks,
defensive rebounds, fouls, offensive rebounds, steals, turnovers, free throw attempts, two point attempts,
three point attempts—all measured per 48 minutes played—plus the percentage of free throw, two-point and
three-point shots made, minutes played, and indicators for missing values. Each of these controls is also
interacted with %White referees. The third column also includes a full set of player × year, home team ×
player race, and team × game fixed effects (including the relevant direct terms).
∗∗∗, ∗∗, and ∗ Statistically significant at 1%, 5%, and 10%.
the racial composition of the refereeing crew is idiosyncratic only
within each season. These control variables are all highly significant,
but nonetheless, the estimated own-race bias is similar to
that estimated in Table III.
Although our player and referee fixed effects take account of
the different styles of individual referees and the different roles
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1870 QUARTERLY JOURNAL OF ECONOMICS
played by individual players, they do not control for how possible
variation in refereeing styles between black and white referees
may differentially impact players with different on-court roles.
The second column addresses this by including a series of controls
for the share of white referees in a game, interacted with variables
describing a player’s on-court role. This set of controls that is interacted
with %white referees (and also included as direct terms)
includes not only all of the controls listed above, but also non–time
varying player characteristics such height, weight, and position;
we also use our sample data to construct measures describing each
player’s on-court role by taking sample averages of each of the
statistics we track (assists, blocks, defensive rebounds, fouls, offensive
rebounds, steals, turnovers, free throw attempts, two-point
attempts, and three-point attempts—all measured per 48 minutes
played—plus free-throw percentage, two-point percentage,
and three-point percentage, minutes played, and indicators for
missing values). Although the full set of these 29 interactions is
jointly statistically significant (although not in the more complete
specification in column (3)), their inclusion does not change our estimate
of the extent of own-race bias. The Online Appendix shows
these interactions, few of which are individually significant. Moreover,
the interaction of %white referees with player race yields
the largest partial and semipartial correlation coefficient of all of
these interactions.
The final column augments this specification with a large
number of fixed effects, which further controls for a range of
competing explanations. This specification includes around 5,000
fixed effects for each player in each year, as well as home team
× player race effects that control for different race effects in each
stadium. Importantly, we saturate the model, allowing for over
25,000 team × game fixed effects (which subsume team × home,
team × year, and team × refereeing crew and many other effects).
These controls ensure that these results are identified only by the
differential propensity of teammates to earn extra fouls when the
refereeing crew is not of their race. Across each of these specifications,
we find that black players receive around 0.18–0.20 more
fouls per 48 minutes played (or 4–41/2%), relative to white players,
when the number of white referees officiating a game increases
from zero to three.
Our dependent variable in these regressions—fouls per 48
minutes—is appropriate if fouls are a linear function of playing
time, which is unlikely given that the six-foul limit is less likely to
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1871
be a constraint for those playing only minor roles. In the extreme
case, a player might be sent into a game with the express purpose
of committing fouls in order to stop the clock in a close game. Thus,
we ran several variants of our baseline regression, finding similar
results when analyzing the foul rate only among starters; controlling
for a quartic in minutes played; or estimating a count model
that includes (log) minutes played as an independent variable.
These results are reported in the Online Appendix.
Table V moves beyond fouls to analyze the consequences of
opposite-race referees for a number of other measurable player
outcomes. Specifically, we measure various box score statistics per
48 minutes played and reestimate equation (1) with that statistic
as the dependent variable. Five main points are evident from this
table. First, we find suggestive evidence of similar effects operating
on flagrant and technical fouls. Although the point estimates
are quite large relative to the rarity of these incidents, they are
also quite imprecise, and only the effect on flagrant fouls is ever
statistically significant. This imprecision reflects the fact that we
have data on these two measures only for 1997/1998–2003/2004,
whereas all other measures are available for the full sample. Despite
the imprecision of these estimates, they are particularly interesting
in that flagrant fouls involve subjective interpretation of
physical contact and technical fouls often involve incidents when
players dispute an on-court ruling.
Second, the propensity to “foul out” appears unaffected by the
race of the refereeing crew, with the 4% rise in the foul rate partly
countered by a 1%–2% decline in playing time. This suggests that
team performance may also be affected by composition effects as
opposite-race referees affect the distribution of playing time.
Third, important effects of own-race bias are evident throughout
the box score. For instance, increasing the share of oppositerace
referees leads to a decline in points scored and a rise
in turnovers committed. The pattern of results across all of
these box score measures—including results that are statistically
insignificant—indicates that player performance appears to deteriorate
at nearly every margin when officiated by a larger
fraction of opposite-race referees. (Note that measured turnovers
include offensive fouls.) Some outcomes may also reflect the role
of the race of the potential “victim” rather than the “offender”
in shaping foul calls. Specifically, these data yield suggestive evidence
of a decline in free throw attempts under opposite-race
referees, suggesting that defensive fouls are less likely to be
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1872 QUARTERLY JOURNAL OF ECONOMICS
TABLE V
EFFECTS OF OPPOSITE-RACE REFEREES ON PLAYER PERFORMANCE
(MEASURED PER 48 MINUTES)
Coefficient on black player ×
% white referees
Dependent Mean
variable (SD) (1) (2) (3)
Personal fouls 4.44 0.197∗∗∗ 0.203∗∗∗ 0.181∗∗
(3.34) (0.061) (0.072) (0.080)
Flagrant fouls 0.012 0.006 0.010∗ 0.009
(0.17) (0.005) (0.006) (0.006)
Technical fouls 0.08 0.007 0.016 0.015
(0.38) (0.010) (0.013) (0.014)
Minutes 30.13 −0.408∗∗∗ −0.503∗∗∗ −0.403∗∗
(10.1) (0.136) (0.160) (0.158)
Fouled out 0.025 −0.000 0.001 0.002
(0.16) (0.003) (0.004) (0.004)
Points 19.54 −0.395∗∗ −0.300 −0.482∗∗
(10.1) (0.176) (0.206) (0.226)
Free throw attempts 5.09 −0.102 −0.018 −0.041
(4.90) (0.090) (0.106) (0.114)
Free throw % 0.75 0.002 0.000 0.001
(0.23) (0.006) (0.007) (0.008)
Blocks 1.02 −0.057∗ −0.011 −0.009
(1.81) (0.030) (0.036) (0.039)
Steals 1.63 −0.062∗ −0.067 −0.078∗
(1.89) (0.036) (0.043) (0.047)
Turnovers 2.95 0.112∗∗ 0.153∗∗∗ 0.121∗
(2.57) (0.050) (0.058) (0.064)
Net effect (win score) 8.36 −0.528∗∗∗ −0.599∗∗∗ −0.509∗∗
(9.09) (0.170) (0.199) (0.218)
Referee, year, and player fixed effects
√ √ √
Player char. × %white referees
√ √
Full set of fixed effects
√
Notes. Each cell reports results from a separate regression. See notes to Table IV for specification details.
Regressions analyzing shooting percentages are weighted by attempts, rather than minutes. n = 266,984,
except flagrant and technical fouls n = 136,509 (available only 1997–2003).
called against one’s opponents when opposite-race players have
possession.
Fourth, the key exception to the general pattern of declining
player performance under opposite-race referees is that a player’s
free throw percentage is unaffected by the racial composition of
the refereeing pool and our estimates on this outcome are quite
precise. We emphasize this result because this is the one on-court
behavior that we expect to be unaffected by referee behavior, thus
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1873
serving as a natural “placebo” measure. Unfortunately field goal
percentage reflects whether the referee assigns blame for physical
contact during the shot to the offense or defense, and hence is not
a useful placebo.5
The final row analyzes a summary measure of a player’s contribution
to his team’s winning margin,6
which suggests that ownrace
bias may lead an individual player’s contribution to his team’s
winning margin to vary by up to half a point per game. Moreover,
the finding that playing time is reduced suggests that there may
be additional consequences due to substitutions.
IV. TEAM-LEVEL ANALYSIS
One shortcoming of our analysis of foul propensities in Table
IV is that it only reflects the role of own-race bias in determining
the guilt of an offender, whereas it may also shape whether a
referee is sympathetic to a player as a victim. By aggregating to
the team level, we can analyze both the number of fouls awarded
against a team, and the number awarded to that team, and see
how these vary with the racial composition of each team and the
refereeing crew. The cost is that aggregating to the team level
substantially reduces the available variation and leads to more
imprecise estimates. Our key estimating equation is
(2)
foulsgrto = β1 %white refereesg × %black minutes playedgt
+ β2 %white refereesg
× opponent %black minutesplayedgo
+ β3 %white refereesg + β4 %black minutes playedgt
+ β5 opponent %black minutes playedgo
+ β6 observable gameg, team–gamegt, and opponent
– gameot characteristics + teamt fixed effects
+ opponento fixed effects + refereer fixed effects
+ seasong fixed effects + observable controls
× %white refereesg + opponent observable controls
5. A score is recorded only if the shooter commits no fouls, whereas a miss is
not recorded if he is awarded a foul.
6. Berri, Schmidt, and Brook (2006) call this index the “Win Score,” and calculate
it as Win Score = (Points − Field goal attempts − 1/2 Free throw attempts) −
Turnovers + Rebounds + Steals + 1/2 Blocks + 1/2 Assists − 1/2 Fouls. We analyze
this productivity index per 48 minutes played.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1874 QUARTERLY JOURNAL OF ECONOMICS
× %white refereesg + %black minutes playedgt
× stadiumg effects + opponent %black minutes playedgo
× stadiumg effects + teamt × seasong effects + opponento
× seasong effects] + εgrto,
where subscript g refers to a particular game, t a particular team,
o their opponent, and r an individual referee. We report standard
errors clustered at the game level.
The extent to which the fouls earned by a team are driven
by their greater racial dissimilarity to the refereeing crew than
their opponents’ is measured by β1 − β2. Note that this estimate
incorporates both the direct effect of the referee’s propensity to
call fouls based on the race of the offender (β1) and the race of the
victim (β2). The net effect on the foul differential (fouls conceded—
fouls awarded) is β1 − β2.
More generally, a shortcoming of the analysis in Table V is
that it only analyzes the effects of refereeing decisions to the extent
that they are captured in individual player box score data.
Indeed, Oliver (2003) notes that a key problem with basketball
statistics is that individual-level box score statistics paint a rich
picture of a player’s offensive production, but they do not reveal
much about either his defensive contribution or general teamwork.
Yet any useful contribution a player makes will be reflected
in the scoring of his team or his opponents, and so we can capture
these contributions by analyzing aggregate team performance.
Consequently we also reestimate equation (2) but analyze points
scored as the dependent variable.
This approach also yields an alternative interpretation that
is particularly useful: changing a team’s racial composition has a
direct effect on the team’s scoring, measured by the coefficient β1
on %white referees × %black minutes played. The same change
in a team’s racial composition also affects its opponents’ expected
scoring, and for the opponent, this effect is measured by β2, the coefficient
on %white referees × %opponent black minutes played.
Thus, β1 measures the effects of own-race bias on a team’s offensive
production, whereas β2 measures the effects on defensive
production, with β1 − β2 measuring the net effect on the winning
margin.
Thus in Table VI we ask whether we see better team
outcomes—fewer fouls committed, more fouls earned, more points
scored, fewer points conceded, and more games won—when a
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1875
TABLE VI
EFFECTS OF OPPOSITE-RACE REFEREES ON TEAM PERFORMANCE
% Black playing time × % white referees
Mean
(SD) (1) (2) (3)
A. Total fouls by team (mean = 22.4)
Total effect 22.4 2.154∗∗ 1.899∗∗ 1.687
(β1 − β2) (4.65) (0.965) (0.940) (1.052)
Of which
Direct effect (β1) 1.135 1.384∗ 1.192
(fouls committed) (0.768) (0.737) (0.817)
Indirect effect (β2) −1.019 −0.515 −0.495
(fouls awarded) (0.793) (0.762) (0.845)
B. Points scored by team (mean = 98.4)
Total effect 98.4 −5.733∗∗∗ −3.836∗∗ −6.185∗∗∗
(β1 − β2) (12.4) (2.011) (1.953) (2.245)
Of which
Direct effect (β1) −2.073 −2.339 −3.202
(points scored) (1.924) (1.792) (2.012)
Indirect effect (β2) 3.660∗ 1.496 2.983
(points conceded) (1.914) (1.800) (2.013)
C. I (home team wins game)
% white refs × −0.195∗∗ −0.160∗ −0.226∗∗
(%blackhome − %blackaway) (0.085) (0.084) (0.092)
% white refs × −0.045 −0.055∗∗ −0.052∗
(black coachhome − black coachaway) (0.028) (0.028) (0.028)
Control variables
Observables; year, referees,
√ √ √
team, and opponent
fixed effects
Full set of fixed effects
√ √
Model OLS OLS IV
Notes. Sample = 24,526 team–game observations in Panels A and B and 12,263 game observations in
Panel C. Each cell reports results from a separate regression. (Standard errors in parentheses, clustered
by game for top two panels.) “Direct” effect refers to coefficient on %black playing time × % white referees.
“Indirect” effect refers to coefficient on opponent %black playing time × % white referees. The total effect is
reported in the top row as the difference.
IV: The endogenous variables %black minutes played, Opponent %black minutes played, and the interaction
of both variables with %white referees are instrumented using Average %black playing time over previous
ten games calculated for both teams, and the interaction of each variable with %white referees. Unreported
“observable” controls include home, attendance, number of overtimes, out-of-contention, and black coach, with
separate control variables recorded for each team.
∗∗∗, ∗∗, and ∗ Statistically significant at 1%, 5%, and 10%.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1876 QUARTERLY JOURNAL OF ECONOMICS
larger fraction of minutes are played by players who are of the
same race as the refereeing crew. Our initial specification includes
observable controls such as whether each team is playing at home,
is out of contention, and has a black coach, game attendance, and
the number of overtimes played; this specification also includes
controls for team, opponent, referee, and season fixed effects. The
full specification also includes the interaction of the observable
control variables with %white referees, as well as separate season
effects for each team, and separate race effects for each stadium;
in each case, each variable is defined for both the team and its
opponent. The number of minutes played by black players may
respond endogenously to the racial composition of the refereeing
crew assigned to a particular game. Consequently, we also
present instrumental variables results in which our variables of
interest—the proportion of each team’s minutes played by blacks,
and that proportion interacted with the racial mix of the referees
on that night—are instrumented with the average share of each
team’s minutes played by black players over that team’s previous
ten games, included both as direct terms and interacted with the
racial mix of the referees on that night. Because team line-ups are
persistent, these are very strong instruments.
For continuity with our earlier analysis, Table VI initially
presents results on the number of fouls awarded against a team.
Although the imprecision in these estimates cautions against a
strong interpretation, we find that the estimated direct effect of
own-race bias on the total number of fouls earned by a team
is roughly five times larger than our estimates of the effect of
own-race bias on the number of fouls earned by an individual
player per 48 minutes. The indirect effect, due to the referee’s
racial similarity to a team’s opponent, is also of a magnitude
roughly similar to that of the direct effect, suggesting that the
analysis of individual data understates the effects of own-race
bias by up to one-half.
Naturally, basketball production is measured not in fouls but
in points scored and conceded. Thus, the second set of results in
Table VI focus on points scored. These estimates again point to a
roughly equal role of own-race bias in shaping a team’s offensive
production as in shaping its defense: the effect of a team’s racial
composition is roughly as large on points scored as it is on the
points scored by its opponent.
The last rows in Table VI examine the effect of racial bias on
whether a team wins. Because one team’s win is its opponents’ loss
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1877
and equation (2) controls symmetrically for the characteristics of
each team, this specification is equivalent to a game fixed-effects
specification or home-versus-away difference regression.7
For simplicity,
we show this equivalent presentation, analyzing whether
the home team won as a function of the home-versus-away difference
in playing time by black players, interacted with the fraction
of white referees, controlling for home–away differences in the
independent variables. These results show quite large and statistically
significant impacts of the mismatch between the racial
composition of the referees and the players.8
In addition, it is
generally believed that coaches have some influence over the decisions
of referees. The bottom row of Panel C provides suggestive
evidence of bias against opposite-race coaches, with the magnitude
of the coach effect being roughly equivalent to the effect of
the race of a single player.
V. QUANTITATIVE INTERPRETATION
The results in Table VI suggest that own-race bias may be an
important factor in determining game outcomes. Figure I provides
a particularly straightforward representation of the data underlying
these findings, plotting local averages of team winning margins
against the proportion of playing time given to black players
relative to the opponents. The slope of these running averages
shows that difference in playing time by black players is correlated
with winning margins. This is not in itself evidence of bias,
as there may be differences in ability. Instead, our analysis highlights
the fact that the slope of this relationship appears to change,
depending upon the racial compositions of the refereeing crew.
It is worth pausing to assess the quantitative importance
of these results and their consistency with our earlier findings.
In order to fix an initial scaling, note that the variable
measuring racial mismatch between players and referees,
(%Blackhome
–%Blackaway
) × %White referees has a standard deviation
of 0.14, suggesting that a one–standard deviation rise in mismatch
reduces a team’s chances of winning by around two to three
7. The home–away difference specification we show yields coefficient estimates
that are exactly half those from estimating equation (2), or the game fixedeffects
specification.
8. Although we report results from a linear probability model, a probit model
yielded similar estimates. For example, whereas the linear probability model in the
first column of Table VI yields a coefficient of −0.196 (with standard error 0.084),
the equivalent probit specification yielded a marginal effect of −0.216 (standard
error 0.091).
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1878 QUARTERLY JOURNAL OF ECONOMICS
–2
–1
0
1
2
–50% –25% 0% 25% 50%
3 white referees
–2
–1
0
1
2
–50% –25% 0% 25% 50%
2 white referees
–2
–1
0
1
2
–50% –25% 0% 25% 50%
1 white referee
–2
–1
0
1
2
–50% –25% 0% 25% 50%
0 white referees
Averagewinningmargin:Pointsforlessagainst
Difference in racial composition of teams: %Black less opponent %black
(measured as difference in share of playing time)
FIGURE I
Effects of Own-Race Bias on Winning Margins
Line shows running mean calculated using Epanechnikov kernel with bandwidth
set to 0.4. Shading shows symmetric 95% confidence intervals (if within
scale).
percentage points. Of course, this one–standard deviation shock
reflects a combination of changes in the racial composition of each
team and changes in the racial composition of the refereeing crew.
We can also use our estimates to assess the sensitivity of
game outcomes to changes in just the racial composition of the
refereeing crew. For instance, in an average game, one team plays
around 15% fewer minutes with black players than their opponent
(which roughly corresponds with that team having one less black
starter). For this team, the chances of victory under an all-black
refereeing crew versus an all-white crew differ by around three
percentage points (= 0.196 × 0.15). Thus, changing the race of just
one referee typically changes the chances of winning by around
one percentage point.
Throughout our sample, the refereeing crew was, on average,
68% white, whereas the teams were 83% black (weighted by
playing time). A different thought experiment considers the consequences
of race-norming the referee pool so that it matches the
racial composition of the player pool. In our sample, the team with
a greater share of playing time accounted for by black players won
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1879
48.6% of their games; our estimates suggest that a race-normed
refereeing panel would raise number by 1.5 percentage points.9
In order to translate these magnitudes into payroll consequences,
consider the following equation from Szymanski (2003),
estimated using team-by-season NBA data from 1986 to 2000:
win percentage,team,year = 0.21 + 0.29
× (team wage billteam,year/league average wage billyear).
Interpreting this as a causal relationship suggests that a 1.5–
percentage point rise in a team’s winning percentage could alternatively
be achieved by raising the aggregate wage bill of an
average team by 1.5%/0.29 ≈ 5%. In turn, consider the modal
game in our sample: a team with five black starters playing four
black starters and one white starter (which occurs in 33% of the
games). The team with the one white starter could maintain its
winning percentage under a shift to race-normed referees either
by upgrading the quality of the team by spending an extra 5% on
player salaries, or by simply exchanging the white starter for a
similar quality black starter. This exercise suggests that the racial
composition of the refereeing pool influences the market value of
white versus black players.
The thought experiment also yields interesting player-level
implications. Given that the large majority of players, on both
the winning and losing sides, are black, race-norming the referee
pool could change a lot of game outcomes but still yield only small
effects on games won by black players (it would rise from 49.8% to
50.0%, as only a few more players would gain than lose). But the
effects on white players would be more dramatic: in our sample,
white starters win around 51.3% of their games; our estimates
suggest that race-norming the refereeing crew would lower this
winning percentage by 1.2 percentage points.
Although these estimates of the number of game outcomes
determined by own-race bias may seem large, a simple example
illustrates that they are consistent with the player-level analysis
in Table V. Consider a game involving five black starters against
four blacks and one white. Any team-level differences will be
9. To see this, note that the average absolute difference in the proportion of
playing time by blacks is around 15%; multiplying this number by the coefficient
of −0.195 yields an estimate of the change in the likelihood of the team with more
minutes played by black players winning the game under an all-white versus
all-black crew. Further scaling by the magnitude of the proposed change in the
proportion of white referees (17%–68%) yields 1.5 percentage points.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1880 QUARTERLY JOURNAL OF ECONOMICS
driven by the differential treatment of the fifth player, who is
black for the home team and white for their rival. Using the
coefficient on Berri, Schmidt, and Brook’s (2006) “Win Score”
metric in Table V, the black player’s overall contribution to the
team’s winning margin will rise by about one-fourth of a point
under a race-normed refereeing crew. These individual-level
estimates are consistent with the estimates of the “direct” effects
measured in Table VI, but that table also shows that these
“direct” effects on fouls committed and points scored are roughly
matched by an equal-sized (and opposite-signed) “indirect”
effect on fouls awarded and points conceded. Consequently
race-norming the refereeing crew would, on average, change the
winning margin by around half a point, which is what we found
in the team-level analysis in Table VI.10
These apparently small
impacts of own-race bias easily yield important effects on win
percentages in a league in which around 61/2% of games go to
overtime, and around 41/2% of game outcomes are determined by
only one point. That is, when game outcomes are typically very
close, even fairly small differences in player performance can
yield large differences in how frequently each team wins.
VI. BEHAVIORAL INTERPRETATION
Thus far our analysis has established that player and team
performance varies with the racial composition of the refereeing
crew. Unfortunately, our framework is not well suited to sorting
out whether these results are driven by the actions of black
or white referees, because this would require establishing a “nodiscrimination”
baseline. Although we can control for enough observable
features of the game so that perhaps our regression
models may establish a reasonable “no-discrimination” benchmark,
it is worth emphasizing that this involves substantially
stronger assumptions than our earlier analysis.
To illustrate this, we analyze our data at the level of the referee.
We use our player–game level data and collect all of the observations
associated with a particular referee. For each referee,
we regress the foul rate against player race, controlling for the
full set of player characteristics noted earlier: height, weight, age,
10. To see this, multiply the points scored regression coefficient in Table VI
(β1β − β2 = 5.733 points) by the difference in playing time given to blacks (20% in
this example), and further multiply by the difference in the share of white referees
(17%–68%), yielding the implication that race-norming referees would lead the
winning margin to change by around half a point.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1881
–0.6
–0.4
–0.2
0
0.2
0.4
0.6
Extrafoulsawardedagainstblackplayers(per48minutes)
Referee-specific estimate of racial bias in foul-calling (sorted)
White referees
Game−weighted ave. = 0.050
Black referees
Game−weighted ave. = −0.009
FIGURE II
Distribution of Racial Bias, by Referee Race
Each point represents an estimate of the number of extra fouls per 48 minutes
an individual referee calls on black versus white players; the bars represent the
95% confidence interval around these estimates. Specifically, we run separate
regressions for each referee, regressing the number of fouls earned per 48 minutes
for each player–game observation in which the referee participated, against an
indicator variable for whether the offending player is black, controlling for year
fixed effects and the full set of player, team–game, and player–game controls and
career statistics listed in the notes to Table IV. All regressions are weighted by
minutes played. The figure only reports results for referees with at least 100 games
in our data set.
experience, all-star status, position, and sample averages of various
box-score statistics (including their usual foul rate). Figure II
plots this estimate for each referee of the degree to which he or
she calls more or fewer fouls on blacks, showing those referees
with at least 100 games in our sample.
This figure illustrates four important features of our analysis.
First, the influence of player race on foul-calling is, on average,
different for white and black referees, with each typically favoring
players of their own race; the magnitude of the difference is consistent
with the estimates reported in Tables III and IV. Second,
there are no individual referees whose racial biases are particularly
notable. (Although a few observations are individually statistically
significantly different from zero, we do not emphasize this
fact, due to the number of referees we test.) Third, the finding of
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1882 QUARTERLY JOURNAL OF ECONOMICS
own-race bias is pervasive across all of our referees: the vast majority
of black referees have a greater propensity to call fouls
against white players than the majority of white referees. Indeed,
despite the imprecision of each referee-specific estimate, only 9
of 28 black referees have an estimated pro-white bias stronger
than the game-weighted average among white referees; similarly,
only 15 of 52 white referees have a weaker pro-white bias than
the game-weighted average among black referees. These findings
suggest that statistically significant evidence of own-race bias persists,
even when our analysis is aggregated to the level of each
individual referee’s record. Fourth, because these regressions are
estimated separately for each referee, they control for referee-byreferee
differences in refereeing “style.”
The simplest interpretation of these results is an own-race
bias on the part of referees. However, there are a few alternative
explanations for our results. First, our results may come from
players changing their behavior in response to the racial mix of
the refereeing crew. Specifically, players would need to play more
aggressively when officiated by more opposite-race referees. However,
although fouls rise under opposite-race crews, Table V yields
no evidence that other measures of aggression, such as steals or
blocks, also rise. Indeed, even if players are unaware of an ownrace
bias by referees, they are aware of their own foul count,
and responding to this alone will yield more careful play under
opposite-race referees. This type of strategic response will lead
to an attenuation bias, making it harder to discern any effects of
own-race bias in the data.
Another possible explanation follows a variant of the usual
“omitted-variables” interpretation of race differences. This alternative
suggests that white and black referees have different focus
areas on the floor, or are trying to penalize different types of behavior.
The omitted variable in this interpretation is the differential
propensity for white or black players to make these types of plays,
and it may be the interaction of different refereeing styles with
different on-court roles that creates the pattern we see in the data.
Some of these possibilities can be addressed by aggregating
to the team level, as in Table VI. For instance, if certain oncourt
roles are typically filled by black players, and these roles
are more harshly penalized by white referees than by black referees,
this would yield a correlation between foul calls and player
race in the individual data. However, aggregating to the team
level aggregates out the differential sorting of blacks and whites
to these roles—particularly if the absence of a black player to
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1883
fill that role would lead to a white player filling it. That is, the
team-level regressions reflect the net impact of changing the racial
composition of playing time, but eliminate variation due to which
players have which roles. The fact that we find roughly consistent
effects in our individual and team-level analyses speaks against
this omitted-variables interpretation.
We also test the sensitivity of our results to various proxies
for the omitted variable by attempting to capture a player’s
“style” through variables measuring his height, weight, age, experience,
all-star status, and position. We also use each player’s
playing history to describe his “style” in terms of the sample average
rates of free-throw attempts, two-point attempts, three-point
attempts, fouls, assists, steals, blocks, turnovers, and offensive
and defensive rebounds earned per 48 minutes played, as well as
free-throw, two-point, and three-point shooting percentages. Interestingly,
these variables do successfully pinpoint an identifiably
black playing style quite successfully—a probit model (not shown)
attempting to predict a player’s race from these “style” variables
yielded a pseudo-R2
of .35, and 11 of 21 variables are individually
statistically significant at a 5% level. Even so, the addition of these
variables to our main regressions (interacted with %white referees
to take account of the different response of white referees to
the different style of black players) does not appreciably change
our estimates of own-race bias (compare columns (1) and (2) of
Table IV). Indeed, these player style × %white referees control
variables are jointly significant only in some specifications, but
are insignificant when controlling for game × team fixed effects.
A third explanation is that black and white referees differ
along a number of dimensions (experience, age, birthplace, etc.)
and it is these differences, rather than race, that explain our results.
For 82% of the games in our sample, we know the NBA
referee experience of all three officials. When we include the average
experience of the crew interacted with the player’s race as
an additional control in our model, the coefficient is both small
and insignificant, and its inclusion has almost no effect on our
estimated own race bias. In addition, for 24% of the games in our
sample, we also know the age of all three referees and how many
of them were born in the South. We interact these additional crewlevel
measures (along with average experience) with each player’s
race and again find that the coefficients on these additional referee
characteristics are small and insignificant, and do not have
a large effect on our estimate of the own-race bias. Full details of
these regressions are provided in the Online Appendix.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1884 QUARTERLY JOURNAL OF ECONOMICS
Our analysis largely proceeds at the player–game level, and
so contrasts the behavior of different refereeing crews, rather than
individual referees. Although this is appropriate in the context of
arbitrary assignment of refereeing crews to games, it admits the
possibility that our findings reflect social interactions within refereeing
crews. That is, perhaps the relative disadvantages conferred
by an increasingly opposite-race refereeing crew reflect referees
exhibiting less own-race bias in the presence of referees not of
their race. In order to isolate the direct influence of individual referees
exhibiting own-race bias from these social interactions, we
reran our analysis of the foul data, focusing only on the contrast
between games refereed by all-black or all-white crews. Comparing
the first and fourth rows of Table III gives a sense of this
analysis, but a more complete analysis—available in the Online
Appendix—shows that even in this restricted set of games we obtain
statistically significant and quantitatively similar estimates
of own-race bias. An alternative regression controls for these crew
composition effects by including dummies for both mixed race
crews, and their interaction with player race; this also yields similar
results to our central findings in Table IV.
As additional support for our main findings, two recent
papers provide evidence of own-race bias of officials in baseball.
In research stimulated by an early draft of this paper, Parsons
et al. (2008) find that a strike is more likely to be called when
the pitcher and umpire are the same race, and Chen (2007)
finds that white umpires provide a larger strike zone to white
pitchers and a smaller strike zone to white batters. These papers
further demonstrate that this own-race bias is influenced by the
amount of monitoring that is in place. In both cases, the own-race
bias completely disappears in stadiums with a QuesTec system
(devices that provide nearly perfect monitoring of the umpire’s
decisions about whether the pitch was a strike). In addition,
Chen (2007) finds that the own-race bias on the part of a white
home-plate umpire is reduced when the umpire works with a
racially diverse crew of officials.
VII. CONCLUSIONS
Using a unique data set on NBA games, we test whether players
of a given race receive fewer fouls when more of the referees
present in the game are of their race. The richness of our data
allows us to control for a host of relevant factors that influence
the number of fouls called and thereby to focus specifically on the
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1885
racial interaction between players and referees. We find that players
have up to 4% fewer fouls called against them and score up
to 21/2% more points on nights in which their race matches that
of the refereeing crew. Player statistics that one might think are
unaffected by referee behavior are uncorrelated with referee race.
The bias in foul-calling is large enough so that the probability of
a team winning is noticeably affected by the racial composition of
the refereeing crew assigned to the game.
These results are striking, given the level of racial equality
achieved along other dimensions in the NBA and the high level of
accountability and monitoring under which the referees operate.
Although the external validity of these results remains an open
question, they are at least suggestive that implicit biases may
play an important role in shaping our evaluation of others, particularly
in split-second, high-pressure decisions. That is, although
these results may be of interest to those intrigued by the sporting
context, we emphasize them instead as potentially suggestive
of similar forces operating in a range of other contexts involving
rapid subjective assessments.
APPENDIX: FURTHER RANDOMIZATION TESTS
Dependent variable: number of white referees
in each game (each cell reports p-values
from F-tests of significance)
Independent variables (1) (2) (3) (4) (5)
Year fixed effects .00 .00 .00 .00 n.a.
#black starters (home) .57 .653 .75 .87
#black starters (away) .41 .40 .72 .42
Attendance .21 .49 .83
Out of contention (home) .98 .94 .60
Out of contention (away) .70 .81 .97
Home team fixed effects .48 .97
Away team fixed effects .97 .71
Home team × year 1.00
fixed effects
Away team × year 1.00
fixed effects
F-test: variables not in .61 .63 .89 1.00
prior column
F-test: all variables except .61 .74 .92 1.00
year effects
Adj. R2 .0495 .0494 .0493 .0483 .0358
Note. Sample includes 12,263 regular-season games.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
1886 QUARTERLY JOURNAL OF ECONOMICS
BRIGHAM YOUNG UNIVERSITY AND NATIONAL BUREAU OF ECONOMIC RESEARCH
WHARTON, UNIVERSITY OF PENNSYLVANIA, BROOKINGS, CENTRE FOR ECONOMIC
AND POLICY RESEARCH, CENTER FOR ECONOMIC STUDIES AND IFO INSTITUTE FOR
ECONOMIC RESEARCH, INSTITUTE FOR THE STUDY OF LABOR, AND NATIONAL BUREAU
OF ECONOMIC RESEARCH
REFERENCES
Abrams, David, Marianne Bertrand, and Sendhil Mullainathan, “Do Judges Vary
in their Treatment of Race?” University of Chicago Mimeo, 2006.
Antonovics, Kate, and Brian Knight, “A New Look at Racial Profiling: Evidence
from the Boston Police Department,” Review of Economics and Statistics, 91
(2009), 163–177.
Berri, David, Martin B. Schmidt, and Stacey L. Brook, The Wages of Wins: Taking
Measure of the Many Myths in Modern Sport (Stanford, CA: Stanford University
Press, 2006).
Bertrand, Marianne, Dolly Chugh, and Sendhil Mullainathan, “Implicit Discrimination,”
American Economic Review, 95 (2005), 94–98.
Bodvarsson, Orn, and Raymond Brastow, “A Test of Employer Discrimination in
the NBA,” Contemporary Economic Policy, 17 (1999), 243–255.
Bushway, Shawn, and Ann Piehl, “Judging Judicial Discretion: Legal Factors and
Racial Discrimination in Sentencing,” Law and Society Review, 35 (2001),
733–764.
Chen, Jowei, “What Does Baseball Teach Us about Reducing Racial Discrimination?
Evidence from Two Natural Experiments,” Stanford University Mimeo,
2007.
Donohue, John, and Steven Levitt, “The Impact of Race on Policing and Arrests,”
Journal of Law and Economics, 44 (2001), 367–394.
Greenwald, Anthony, and Mahzarin Banaji, “Implicit Social Cognition: Attitudes,
Self-Esteem, and Stereotypes,” Psychological Review, 102 (1995), 4–
27.
Hamilton, Barton, “Racial Discrimination and Professional Basketball Salaries in
the 1990s,” Applied Economics, 29 (1997), 287–296.
Kahn, Lawrence, “Discrimination in Professional Sports: A Survey of the
Literature,” Industrial and Labor Relations Review, 44 (1991), 395–
418.
Kahn, Lawrence, and Malav Shah, “Race, Compensation and Contract Length in
the NBA: 2001–2,” Industrial Relations, 44 (2005), 444–462.
Kahn, Lawrence, and Peter Sherer, “Racial Differences in Professional Basketball
Players’ Compensation,” Journal of Labor Economics, 6 (1988), 40–61.
Koch, James, and C. Warren Vander Hill, “Is There Discrimination in the ‘Black
Man’s Game’?” Social Science Quarterly, 69 (1988), 83–94.
Oliver, Dean, Basketball on Paper: Rules and Tools for Performance Analysis
(Washington, DC: Potomac Books, 2003).
Parsons, Christopher, Johan Sulaeman, Michael Yates, and Daniel Hamermesh,
“Strike Three: Umpires’ Demand for Discrimination,” IZA Discussion Paper
No. 3899, 2008.
Payne, Keith, Alan Lambert, and Larry Jacoby, “Best Laid Plans: Effects of Goals
on Accessibility Bias and Cognitive Control in Race-Based Misperceptions
of Weapons,” Journal of Experimental Social Psychology, 38 (2002), 384–
396.
Schanzenbach, Max, “Racial and Sex Disparities in Prison Sentences: The Effect
of District-Level Judicial Demographics,” Journal of Legal Studies, 34 (2005),
57–92.
Spohn, Cassia, “The Sentencing Decisions of Black and White Judges: Expected
and Unexpected Similarities,” Law and Society Review, 24 (1990), 1197–
1216.
Stauffer, Joseph M., and M. Ronald Buckley, “The Existence and Nature of Racial
Bias in Supervisory Ratings,” Journal of Applied Psychology, 90 (2005), 586–
591.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom
RACIAL DISCRIMINATION AMONG NBA REFEREES 1887
Stoll, Michael, Steven Raphael, and Harry Holzer, “Black Job Applicants and the
Hiring Officer’s Race,” Industrial and Labor Relations Review, 57 (2004), 267–
287.
Szymanski, Stefan, “The Economic Design of Sporting Contests,” Journal of
Economic Literature, 41 (2003), 1137–1187.
Timmerman, Thomas, “Racial Diversity, Age Diversity, Interdependence, and
Team Performance,” Small Group Research, 31 (2000), 592–606.
Welch, Susan, Michael Combs, and John Gruhl, “Do Black Judges Make a Difference?”
American Journal of Political Science, 32 (1988), 126–136.
atIowaStateUniversityonAugust17,2011qje.oxfordjournals.orgDownloadedfrom