Race & Quarterback Survival

Article
Journal of Sports Economics

2017, Vol. 18(8) 850-866
ª The Author(s) 2015
Race and Quarterback Reprints and permission:
sagepub.com/journalsPermissions.nav
Survival in the National DOI: 10.1177/1527002515609659
journals.sagepub.com/home/jse
Football League
Brian D. Volz1
Abstract
This study examines data from the 2001 to 2009 National Football League (NFL)
seasons to determine whether Black quarterbacks face discrimination. When con-
trolling for injury, age, experience, performance, team investment, backup quality,
and bye weeks, Black quarterbacks are found to be 1.98–2.46 times more likely to be
benched. Marginal evidence is also found that Black quarterbacks face less dis-
crimination in areas with a larger percentage of Black residents. Additionally, it has
been observed that when White quarterbacks are benched, the team improves by
more than when Black quarterbacks are benched. This provides evidence that there
is a cost to this discrimination.
Keywords
race, NFL, survival analysis, quarterback
Introduction
From 2001 to 2014, the number of Black opening day quarterbacks in the National
Football League (NFL) has varied from as low as 5 to as high as 9. This represents
only 16–28% of starting quarterbacks. These percentages are very surprising consid-
ering that Black athletes account for approximately 67% of the NFL. Black players
also appear to dominate other offensive positions. For example, Black players
1
Department of Economics and Global Studies, Assumption College, Worcester, MA, USA
Corresponding Author:
Brian D. Volz, Department of Economics and Global Studies, Assumption College, 500 Salisbury St.,
Worcester, MA 01609, USA.
Email: bd.volz@assumption.edu
mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Volz 851
represented 83% of running backs and 84% of receivers in 2013 (Lapchick, Dono-
van, Rogers, & Johnson, 2014). This discrepancy leads one to ask why there are not
more Black starting quarterbacks in the NFL.
One possible explanation is that Black quarterbacks face some form of discrim-
ination in the NFL labor market. One way to determine whether Black quarterbacks
are facing discrimination is to analyze the hiring decisions of firms. However, this
approach presents several practical difficulties. For example, it may be very difficult
to identify the pool of all players who would like to be NFL quarterbacks. Addition-
ally, the major qualification for NFL quarterbacking is experience in college foot-
ball. Given the vast number of college football teams which compete in different
leagues and at different levels, comparisons of college statistics between players
would prove problematic at best.
An alternative approach is to consider whether the Black starting quarterbacks
already in the league are treated differently than comparable White quarterbacks.
One way to address this issue is to compare the salaries of White and Black players.
This is the approach that most previous studies of discrimination against NFL play-
ers have taken. The earliest such studies, Mogull (1973, 1981), find no evidence of
salary discrimination against Black players. Kahn (1992) examines salary data from
the 1989 season and finds limited evidence of salary discrimination against Black
players in metropolitan areas with relatively White populations. However, Kahn
notes that the magnitude of this discrimination is less than 4% and insignificant in
most equations. Based on salary data from the 1996 season, Gius and Johnson
(2000) find evidence that Black players are actually paid 10% more than White play-
ers, suggesting reverse discrimination. Berri and Simmons (2009) also analyze player
salaries, specifically looking at the quarterback position from 1995 to 2006. They find
evidence of salary discrimination against Black quarterbacks in the upper half of the
wage distribution. Additionally, they find that Black quarterbacks are more likely to
run but do not appear to be rewarded for that skill. Most recently, Keefer (2013) finds
evidence of salary discrimination against Black linebackers from 2001 to 2009.
Unlike most previous studies, the models presented here focus on survival rather
than salaries. If Black players face employer or customer discrimination, it is likely
that they will be treated differently in more than just compensation. Specifically, in
the presence of discrimination, Black players may be shown less patience by the coa-
ches and fans. This lack of patience may lead to young Black players being given
less opportunity to develop and experienced Black players being benched more
quickly than comparable White players. In other words, the presence of discrimina-
tion against Black players will make them less likely to survive from one season to
the next and from one game to the next than comparable White players. Survival
from season to season has previously been analyzed by Conlin and Emerson
(2006). They examine the impact of race on whether drafted players remain under
contract during their first three seasons. They find that non-White players are actu-
ally more likely to survive than observationally equivalent White players who were
drafted in the same position. They interpret this finding as evidence that non-White

852 Journal of Sports Economics 18(8)
players face discrimination in hiring but not retention. The analysis presented here
further expands the literature on discrimination in professional football by consider-
ing survival from game to game rather than season to season.
The models presented here apply survival time analysis to data on starting quar-
terbacks during the 2001–2009 seasons to determine whether evidence exists that
Black quarterbacks are being treated differently than comparable White quarter-
backs. By focusing on survival rather than salaries, the analysis presented here
approaches the issue of discrimination in a different manner than previous studies.
While survival analysis has not been applied to NFL quarterbacks, it has been used
previously to examine the role of race in sports. Volz (2009) uses survival analysis to
evaluate the role of race in the survival of Major League Baseball managers. This
article aims to contribute to the ongoing debate over the existence of discrimination
in professional football by applying these survival analysis techniques to the NFL.
Variable and Sample Selection
The goal of this analysis is to determine whether race has a significant impact on the
probability of a starting quarterback surviving from one week to the next. Therefore,
the sample consists of streaks of games started by quarterbacks. It is possible that a
starting quarterback is not actually being considered for the starting job the following
week. For example, a quarterback may be filling in for an injured player on a temporary
basis. This is much less likely to be the case for a quarterback starting the first game of
the season than a quarterback who begins their streak in the middle of the season. There-
fore, for each season, a streak is constructed for each week one starting quarterback.
The streak continues until that quarterback does not start the next week or reaches the
end of the regular season. These streaks were constructed for the 2001-2009
seasons.1 This results in 287 streaks comprised of 2,865 observations.2
The central interest of this analysis is the effect of being Black on the likelihood
of survival. Therefore, a variable which is equal to 1 for Black quarterbacks and 0
otherwise is included as a covariate in the analysis. This results in 522 Black quarter-
back observation. One would expect that a Black quarterback is less likely to be dis-
criminated against by a Black coach. In order to capture this effect, an interaction
between the binary variables Black and Black Coach is included. Additionally, one
would expect a Black quarterback to face less discrimination from Black fans.
Therefore, an interaction between the binary variable Black and the percentage of
Black residents in each team’s metropolitan area is also included.3
In addition to race, there are several other variables, which are expected to influ-
ence quarterback survival. The most obvious reason a quarterback would not start
the following week is if they are injured. Fortunately, NFL teams release injury
reports each week which list all injured players that week along with their status for
the upcoming game. While these reports are rather subjective, generally players
listed as out or doubtful on the report do not play that week. Therefore, a variable

Volz 853
equal to 1 for any player who is listed as out or doubtful and 0 otherwise is included
in the analysis to control for injuries.4
As a quarterback ages, they may gain life experience and football-specific skills
that make them a better starting quarterback. Therefore, older quarterbacks may be
more likely to survive from one game to the next. However, teams may also have
more patience with young quarterbacks who are still learning the game. Therefore,
young quarterbacks may be more likely to survive from one game to the next. In
order to control for these effects of age and to determine whether age increases or
decreases survival, a variable equal to each players’ age in years is included in the
analysis. In many circumstances, age may have a diminishing effect. Therefore, the
square of age is also included in the model.
It is also expected that experience as a starting quarterback will increase the like-
lihood of survival. As they play more games, quarterbacks should become more famil-
iar with their team’s plays and players. Experienced quarterbacks may also have more
familiarity with their opponents which could translate into better decision making.
More experience as a starting quarterback also translates into more experience dealing
with the media. A quarterback who ‘‘says all the right things’’ may be able to survive a
poor performance better than an inexperienced quarterback. Based on these ideas, a
variable which is equal to the number of games a quarterback has started in their career
is included in the analysis. Like age, the gains from experience may diminish over
time, and therefore the square of games started is also included in the model.
In addition to his experience on and off the field, a quarterback’s survival should
depend on his level of performance on the field. Decisions on whether to bench a
quarterback or not are made on a game-by-game basis, and therefore, it is expected
that performance in the most recent game will have a significant impact on the like-
lihood of being benched. While recent performance is important, it is unlikely that a
coach will make a decision based on one game alone. Therefore, a player’s career
statistics may also be important. In order to capture the impact of recent and past
performance, a variety of performance statistics from recent performances and a
player’s career are included as covariates.
Most quarterbacks are valued on the field primarily for their ability to complete
passes. Therefore, a measure of passing ability must be included as a covariate. The tra-
ditional measure of passing efficiency is the quarterback rating. This statistic combines
information on completion percentage, yards gained per attempt, touchdown passes per
attempt, and interceptions per attempt, into a single measure with a maximum value of
158.3. Quarterback ratings above 100 are considered exceptional performances. The
primary advantage of this measure is that it compares quarterbacks to an ideal mea-
sure in each of the four categories so that no one category is weighted more
heavily than another. The quarterback rating for each individual game and for
the quarterback’s career are calculated at the end of each game and included
as covariates in determining the likelihood of surviving to the next game.
While passing is the traditional role of a quarterback, some quarterbacks are
also valued for their ability to run the ball. Therefore, a measure of rushing

performance must also be included as a covariate. Unfortunately, a widely

accepted single measure of rushing ability does not exist. As a result, two rushing
statistics are included in the analysis. These measures include yards per attempt
and touchdowns per attempt. Per attempt measures are preferred to total yards
or touchdowns, as some quarterbacks are asked to run more often than others.
Therefore, having more total yards or touchdowns would not necessarily imply
a better rushing performance.
While individual performance is important, it can be argued that the most impor-
tant thing to fans and coaches is team performance. Therefore, whether the quarter-
back wins each week is also included as a covariate. Teams are likely more
concerned about their league standing than whether they win or lose any one specific
game. Therefore, a quarterback’s record over the entire season may also be an
important factor in his survival. In order to capture this effect, a variable equal to
the quarterback’s winning percentage up to that point in the season is included in the
analysis. It is certainly possible that a team could lose, despite an exceptional per-
formance by their quarterback. If the team loses, despite the offense scoring a large
number of points, then teams may not blame the quarterback for the loss. Therefore,
the number of points the offense scores each week is also included in the model.
In addition to evaluating a player’s performance, teams may also take into
account how much they have invested in a specific player. For example, if a team
is paying a quarterback a large salary, they may be less willing to bench that quarter-
back. This may be due to the fact that they would like to avoid the embarrassment of
having made a wrong decision. The NFL also operates under a salary cap system.
The more money a team pays to one quarterback, the less money they have available
to pay another. Therefore, teams may be essentially stuck with a quarterback if they
spent a large portion of the salary cap on that player. Additionally, previous research
by Keefer (2015) finds evidence that an increase in a rookie’s salary cap value
increases their number of games started. Therefore, in order to control for the effect
of a player’s salary on their survival, the percentage of salary cap going toward each
quarterback is calculated for each season and included as a covariate.5
As mentioned earlier, the quality of the backup quarterback may also influence a
team’s decision to bench their starting quarterback. In order to control for this effect,
previous games started by the backup quarterback for each team are included as a
covariate. Experience as a starting quarterback is chosen over performance mea-
sures, as many backup quarterbacks have very little playing experience. Therefore,
their performance measures may be misleading. Switching from one quarterback to
another may be a difficult process, especially if the backup quarterback does not
have significant experience. Therefore, teams may be more likely to switch to the
backup quarterback if they have extra time to prepare. This would be the case if the
team has a bye week. Therefore, a variable equal to 1 if the team has a bye the fol-
lowing week and 0 otherwise is included in the analysis.
Similar to having money invested in a player, teams may also have a high draft
pick invested in a player. Teams may be more patient with high draft picks, as they

Volz 855
have higher expectations for them. Management may also want to avoid the embar-
rassment of having wasted a high draft pick on a player who is not playing. There-
fore, each player’s draft position is also included as a covariate.6
Taken together, these variables control for race, injury, age, experience, individ-
ual performance, team performance, team investment, backup quality, and bye
weeks. Having compiled these statistics for each week one starting quarterback from
2001 to 2009, the resulting 287 streaks are analyzed using survival time analysis as
described in the following section.
Methodology
The goal of this analysis is to determine whether race has a significant impact on the
probability of a starting quarterback surviving from one week to the next. This can
be analyzed using the technique of survival time analysis. The goal of survival anal-
ysis is to estimate a survival function which gives the probability of survival to a
certain time period, given a set of covariates. Survival analysis can be conducted
by making distributional assumptions about the survival function. Models that make
such assumptions are referred to as parametric models. These models estimate the
hazard rate as a function of the covariates. The hazard rate is simply the dropout rate
in a given time period conditional on a set of covariates. This rate is always positive,
so the model is assumed to be linear in the log of the hazard rate. Therefore, the
model of interest is the following:
logðhi Þ ¼ B0 þ B1 Xi1 þ B2 Xi2 þ ::: þ Bn Xin :
This model is independent of time and leads to a survival function of the follow-
ing form:
SðtÞ ¼ eht :
This model, referred to as the exponential survival model, is the most sim-
plistic of the parametric survival models because of its assumption of a constant
hazard rate over time. While this model is valued for its simplicity, the assump-
tion that the hazard rate is constant over time is often inappropriate, and there-
fore models which allow for the hazard rate to vary over time may be more
appropriate. The most common of these parametric models are the Weibull and
Gompertz models. These models assume that there is some underlying hazard
rate which varies with time. It is also assumed that there is no interaction
between time and the covariates. The covariates effect the hazard rate by pro-
portionally changing the underlying rate for a given time period. This is why
these models are referred to as parametric proportional hazards models. Specif-
ically, the hazard rate is modeled as follows:
hðtÞ ¼ h0 ðtÞexpðB0 þ B1 Xi1 þ B2 Xi2 þ ::: þ Bn Xin Þ:

Different functional forms of h0(t) will lead to different survival functions. The
underlying hazard rate is commonly assumed to have a Weibull distribution. This
leads to the following survival function:
SðtÞ ¼ eðhtÞp :
This survival function is useful as the hazard rate will be either increasing or
decreasing monotonically with time depending on the value of the estimated para-
meter p. If p is greater than 1, the hazard rate is increasing over time. If p is less than
1, the hazard rate is decreasing over time. If p is equal to 1, this implies the hazard
rate is independent of time and leads to the exponential model previously presented.
Another commonly used distributional assumption is the Gompertz distribution.
If the underlying hazard is assumed to follow a Gompertz distribution, the resulting
survival function is of the following form:
SðtÞ ¼ exp½ðh=rÞð1 ert Þ:
Under this assumption, the hazard rate will either increase or decrease at an expo-
nential rate depending on the value of the estimated parameter r. Taken as a group,
these three distributional assumptions can be used to estimate a hazard rate which is
constant over time, increasing over time, or increasing exponentially over time.
These three parametric models are estimated using maximum likelihood. The results
are presented in Tables 1, 2, and 3 and are discussed in the next section.
An alternative approach to these parametric models is to estimate a Cox propor-
tional hazards model. To estimate this model, the ratio of hazards for two observa-
tions is taken as follows:
hi ðtÞ=hj ðtÞ ¼ h0 ðtÞexpðB0 þ B1 Xi1 þ :: : þ Bn Xin Þ
=h0 ðtÞexpðB0 þ B1 Xj1 þ :: : þ Bn Xjn Þ:
Because of the fact that the baseline hazards are independent of the covariates, the
baseline hazards cancel leaving the following hazard ratio which is independent of
time:
hi ðtÞ=hj ðtÞ ¼ expðB0 þ B1 Xi1 þ ::: þ Bn Xin þÞ=expðB0 þ B1 Xj1 þ ::: þ Bn Xjn Þ:
Despite the fact that the underlying hazard function is not defined, the model can
still be estimated by the method of partial likelihood. This partial likelihood method
is presented by Cox in the 1972 article in which he first introduces the Cox model
(Cox, 1972). Although these models are not as efficient as a correctly specified para-
metric model, they do not depend on distributional assumptions. This avoids the risk
of obtaining misleading results because of an incorrectly specified parametric
model. A Cox proportional hazards model is also estimated with the results pre-
sented in Table 4 and discussed in the next section.
The advantage of the Cox model is that it does not rely on distributional assumptions
about the survival function. However, the model is only appropriate if the assumption of

Volz 857
Table 1. Exponential Model.
All Variables Included Significant Variables
Covariate Hazard Ratio P-Value Hazard Ratio P-Value
Black 2.20 .057* 1.42 .041**

Black Metro % Black 0.98 .196 —
Black Black Coach 0.84 .720 —
Injury List 15.95 .000*** 16.45 .000***
Age 1.15 .607 1.03 .026**
Age2 1.00 .699 —
Career Starts 1.00 .925 —
Career Starts2 1.00 .811 —
Salary Cap % 0.00 .007*** 0.00 .000***
Draft Pick 1.00 .556 —
Bye Next Week 0.58 .072* 0.62 .076*
Backup QB Experience 1.00 .024** 1.00 .015**
Win 0.70 .115 0.70 .087*
Winning Percentage 0.41 .003*** 0.38 .001***
Offensive Points 0.98 .025** 0.97 .003***
Passer Rating 1.00 .552 —
Rush Yards/Attempt 0.99 .460 —
Rush TD/Attempt 1.28 .685 —
Career Passer Rating 0.99 .281 —
Career Rush Yards/Attempt 1.03 .624 —
Career Rush TD/Attempt 1.34 .854 —
Note. Streaks ¼ 287, n ¼ 2,865, TD ¼ Touchdowns, Wald test of coefficients equal to 0: p ¼ .0000.
*Significant at the 10% level. **Significant at the 5% level. ***Significant at the 1% level.
proportional hazards is satisfied. The proportional hazards assumption can be tested by

running a generalized linear regression of the scaled Schoenfeld residuals on time.
A significant coefficient on time would provide evidence that the proportional
hazards assumption is violated. A global test for the time covariates fails to reject
the null hypothesis of zero slope for all models estimated. Therefore, the Cox pro-
portional hazards model is appropriate to use in this analysis. Wald tests of whether
the coefficients are equal to 0 are also rejected with p-values of .0000 for all models
estimated.
Results
Tables 1–3 present the estimated hazard ratios from the three parametric survival
models. Table 4 presents the results from the Cox Proportional Hazards model. A
hazard ratio greater than 1 implies that the covariate decreases the likelihood of sur-
vival, while a hazard ratio less than 1 implies that the covariate increases the

Table 2. Weibull Model.
Black 2.76 .025** 2.59 .014**

Black Metro % Black 0.97 .119 0.97 .102
Injury List 15.98 .000*** 16.34 .000***
Age 1.22 .508 1.03 .037**
Age2 1.00 .650 —
Salary Cap % 0.00 .033** 0.00 .000***
Bye Next Week 0.58 .071* 0.60 .074*
Win 0.77 .205 —
p 1.37 1.35
Note. Streaks ¼ 287, n ¼ 2,865, TD ¼ Touchdowns, Wald test of coefficients equal to 0: p ¼ .0000.
likelihood of survival. In addition to the model with all variables included, each table
also presents a model with only statistically significant variables included. These
models were selected by eliminating one at a time the least significant variables until
all variables were significant at the 10% level. The only variable which goes from
highly insignificant to significant when other variables are eliminated is age. This
is likely due to the correlation between Age and Age2. None of the other covariates’
p-values change meaningfully with the elimination of other variables. In addition to
showing that the results are robust to variable selection, this provides some evidence
that multicollinearity is not a problem in this model.
The variable of interest in this analysis is Black. As can be seen in Tables 1–4, the
hazard ratio for Black is greater than 1 for all model specifications. This result is sta-
tistically significant with p-values ranging from .005 to .057 depending on the
model. A statistically significant hazard ratio which is greater than 1 implies that
Black quarterbacks are less likely to survive to the next week than observationally
equivalent White quarterbacks. Additionally, the hazard ratio for Black Metro

Volz 859
Table 3. Gompertz Model.
Black 2.56 .032** 1.43 .044**

Black Metro % Black 0.97 .131 —
Injury List 16.62 .000*** 16.72 .000***
Age 1.21 .504 1.03 .034**
Age2 1.00 .641 —
Salary Cap % 0.00 .023** 0.00 .000***
Bye Next Week 0.62 .116 —
Win 0.73 .142 0.71 .092*
r 0.06 0.06
Note. Streaks ¼ 287, n ¼ 2865, TD ¼ Touchdowns, Wald test of coefficients equal to 0: p ¼ .0000.
% Black is less than 1 and statistically significant at the 10% level for the Cox model.
This provides limited evidence that Black quarterbacks face less discrimination in
areas with a higher percentage of Black residents. It should be noted, however, that
the overall marginal effect of being Black remains negative for all teams. The hazard
ratio for Black Black Coach is found to be statistically insignificant for all model
specifications. This implies that Black quarterbacks with Black coaches are not
treated differently than those with White coaches. However, it should be noted that
there is a very small number of instances where a Black quarterback played for a
Black coach in this sample. Therefore, this insignificant result may simply reflect
a lack of variation in the data.
Figure 1 presents predicted survival functions for each model specification with
all variables included. The predicted survival functions for a White quarterback ver-
sus a Black quarterback are plotted with all other covariates evaluated at their means.
As is evidenced by the lower curves for Black quarterbacks, Black quarterbacks
have a lower probability of survival each week. Specifically, Black quarterbacks are

Table 4. Cox Proportional Hazards Model.
Black 2.82 .018** 2.88 .005***

Black Metro % Black 0.97 .098* 0.97 .068*
Injury List 16.20 .000*** 16.71 .000***
Age 1.24 .442 1.04 .017**
Age2 1.00 .532 —
Salary Cap % 0.00 .010*** 0.00 .000***
Bye Next Week 0.62 .112 —
Backup QB Experience 1.01 .002*** 1.01 .001***
Win 0.70 .100* 0.69 .070*
Note. Streaks ¼ 287, n ¼ 2865, TD ¼ Touchdowns, Wald test of coefficients equal to 0: p ¼ .0000.
1.98 to 2.46 times more likely to be benched than an observationally equivalent

White quarterback depending on the model and week chosen. The predicted prob-
abilities of being benched for the Cox model are presented in Table 5. For example,
in Week 8 of the season, the Cox model predicts that an average White quarterback
has a 4% chance of being benched, while an average Black quarterback has a 9.7%
chance of being benched. As can be seen in Table 5, the model estimates that Black
quarterbacks are more than twice as likely to be benched as White quarterbacks for
each week. This provides evidence that Black quarterbacks may face some form of
discrimination in the National Football League. This evidence appears robust to
model specification and variable selection.
According to Becker (1957) discrimination may be based on the preferences of
employees, customers, or employers. An implication of Becker’s model of employee dis-
crimination is that in the presence of such discrimination we should see segregated work
forces. Given that the NFL is far from segregated, it is unlikely that any discrimination
against Black quarterbacks comes from fellow players. This seems especially unlikely
considering that the vast majority of the receivers quarterbacks throw to are Black.

Volz 861
Exponential Model Weibull Model

1
1 .8
.8
Survival
Survival
.6
.6
.4
.4
.2
0 5 10 15 0 5 10 15
Week Week
Black White Black White
Gompertz Model Cox Proportional Hazards Model

1
1 .8
.8
Survival
Survival
.6
.6
.4
.4
.2
0 5 10 15 0 5 10 15
Week Week
Black White Black White
Figure 1. Estimated survival functions by race.
Table 5. Predicted Probability of Being Benched.
Cox Model Including All Variables
Week White Black
1 0.030 0.074
2 0.016 0.038
3 0.053 0.127
4 0.035 0.086
5 0.018 0.044
6 0.054 0.128
7 0.028 0.068
8 0.040 0.097
9 0.048 0.115
10 0.034 0.082
11 0.035 0.086
12 0.031 0.074
13 0.038 0.092
14 0.019 0.047
15 0.039 0.095

Table 6. Difference in PPG Between Replacement and QB.
QBs not Appearing on Injury Report 2001–2009
Mean N Standard deviation
White 1.756 235 9.027

Black 0.755 56 9.559
Note. Test of Black–White <0: p ¼ .038. PPG ¼ Points Per Game, QB ¼ Quarterback.
Alternatively quarterbacks may be facing customer discrimination. A major

implication of the customer discrimination model is that firms may attempt to hide
minority labor from their customers. On most NFL teams, the quarterback is the
most visible player both on and off the field. It is possible that teams may be
responding to customer preferences by avoiding Black quarterbacks and attempting
to hide Black labor at less visible positions.
It is also possible that Black quarterbacks are facing employer discrimination. In
his analysis of employer discrimination, Becker argues that discriminating employ-
ers or firms should be driven from a competitive labor market by firms which are
willing to take advantage of minority labor. However, with only 32 teams and very
limited entry of new teams, it can be argued that the NFL is not a competitive mar-
ket. Therefore, it may be possible for discriminating employers to operate in the NFL
without being driven from the market.
If this is the case, one would expect to see evidence of the costs of discrimination
in terms of team performance. In theory, a team should only be willing to replace
their starting quarterback if there is some expected gain to doing so. This will only
be the case when the expected points with the backup are greater than the expected
points with the starter. However, if the team exercises taste discrimination as
described in Becker (1957) or has asymmetric beliefs about Black and White quar-
terbacks as presented in Arrow (1973), they may act as if a Black quarterback has
lower expected points than they actually do. This may cause them to bench a Black
starter when the expected points for the backup are equal to or even less than that of
the starter. If teams are making decisions in this manner then on average the payoff
when replacing a Black quarterback will be less than the payoff when replacing a
White quarterback. In other words, the presence of a Black starting quarterback will
cause teams to make worse decisions.
In order to test whether this is the case, the average points per game are calculated
for each starting quarterback and his replacement for all quarterbacks who were
benched during the sample period. The original quarterback’s points per game are
then compared to the average points per game achieved by the quarterback replacing
them. The average differences between the starting quarterback and replacement
quarterback by race are presented in Table 6. In order to avoid replacements due
to injury, only those quarterbacks who were not listed on the injury report when

Volz 863
benched are included. As can be seen in Table 6, on average, when a White quarter-
back is replaced the team scores 1.76 points more per game under the new quarter-
back. However, on average, when a Black quarterback is replaced, the team scores
.75 points less per game under the new quarterback. This provides evidence that
teams do indeed make worse decisions when replacing Black quarterbacks. This
difference of 2.5 points per game is statistically significant with a p-value of .038.
During the 2014 NFL season, 9% of games were decided by two points or less and
20% of games were decided by three points or less. Therefore, it is likely that over a
16-game season discriminating teams may lose an additional game or two because of
their poor decision making. This provides evidence that discriminating against Black
quarterbacks does have a meaningful cost. However, whether teams are willing to
bear this cost based on their own tastes, as a result of asymmetric beliefs, or in
response to the preferences of their customers remains unclear.
When considering performance measures, it appears that team performance is
more important than individual performance. This is evidenced by a statistically
significant hazard ratio of less than 1 for Winning Percentage for all model specifi-
cations. This ratio implies that having a high winning percentage in the current sea-
son increases a quarterback’s likelihood of starting the next game. This is not
surprising as winning games should be the focus of every football coach when mak-
ing personnel decisions. The hazard ratio on Win is only marginally significant in
some model specifications. These two observations combined imply that teams are
more concerned with their overall league standing than what has happened in the
most recent game. This makes sense as it is the overall win percentage that deter-
mines whether a team makes the playoffs, not what has happened in the most recent
game. Additionally, Offensive Points is statistically significant in all models with a
hazard ratio less than 1. This implies that the more points a team scores the less
likely the quarterback is to be benched independent of how those points are scored.
This implies that a poor performance by a quarterback may be overlooked if the
offense is still able to score an appropriate level of points.
Unlike team performance, it appears that individual performance measures do not
have a statistically significant impact on quarterback survival. The hazard ratios on
all passing and rushing measures are statistically insignificant for all model specifi-
cations.7 This combined with the significant results for team performance measures
implies that quarterbacks are judged based on the performance of the team and
offense as a whole. It appears that quarterbacks are held responsible for poor team
or offensive performance independent of their own performance statistics.
If Black quarterbacks are being discriminated against, it may not only be the
case that they have a lower overall probability of survival but also that their per-
formances are evaluated differently than those of White quarterbacks. In order
to test whether individual and team performances have a different impact on
survival for Black quarterbacks, interaction terms between Black and the perfor-
mance variables were also added to the models. None of these interaction terms
were significant at conventional levels. This implies that individual or team

performance does not marginally impact Black quarterbacks any differently than
White quarterbacks.
The Age and Age2 variables are individually and jointly insignificant for the
models which include both variables. However, Age becomes significant when Age2
is eliminated from the model. A significant hazard ratio greater than 1 for Age
implies that younger quarterbacks are more likely to survive than older quarterbacks.
This may reflect the idea that teams have more patience with young quarterbacks
who are learning than they do with older quarterbacks.
It appears that higher paid starting quarterbacks are more likely to survive from
one week to the next. This is evidenced by the hazard ratios being less than 1 and
statistically significant for the Salary variable under all models and samples. This
implies that teams may be less willing to bench a quarterback in whom they have
a large financial investment. Unlike their salary, it appears that a player’s draft pick
does not influence his likelihood of survival. The variable Draft Pick is insignificant
for all models and samples.
Conclusion
This analysis has shown that when controlling for injury, age, experience, individual
performance, team performance, team investment, backup quality, and bye weeks
Black starting quarterbacks are 1.98 to 2.46 times more likely to be benched the next
game than observationally equivalent White quarterbacks. This result is statistically
significant and appears robust to model and variable selection. This implies that
Black quarterbacks may face some level of discrimination in the NFL. Limited evi-
dence is found that Black quarterbacks face less discrimination in areas with a larger
percentage of Black residents. This implies that at least some of this discrimination
may be attributable to the customers. Additionally, it has been observed that when
White quarterbacks are replaced, the team improves by more than when Black quar-
terbacks are replaced. This provides evidence that there is a cost to this discrimina-
tion. Given that the NFL is not a perfectly competitive market it is possible that
discriminating employers are not being driven out of the market as economic theory
would predict. However, it is also possible that owners are simply responding to the
desires of their customers. Future research should seek to better analyze the sources
of this discrimination by identifying situations where employer or customer prefer-
ences can be observed independently.
Acknowledgments
The author would like to acknowledge the contributions of the editor and two anonymous
referees whose comments greatly improved the quality of this research. The author would also
like to thank participants at the 2012 Eastern Economic Association’s Annual Conference for
their helpful comments on an earlier version of this article. As always any mistakes are the
author’s own.

Volz 865
Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, author-
ship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of
this article.
Notes
1. The 2010 National Football League (NFL) season was played without a salary cap. There-
fore, the chosen sample ends in 2009 to avoid including a year with an inconsistent salary
measure.
2. Team performance and individual performance and characteristics are based on publicly
available data from http://www.pro-football-reference.com.
3. These data were taken from the 2009 Census Bureau State and Metropolitan Area Data
Book, table B-3.
4. Models were also estimated using separate binary variables for each possible listing on the
injury report. The results from these models did not meaningfully differ from those pre-
sented here.
5. Player salary cap values are based on the USA Today NFL Salary Database available at
http://content.usatoday.com/sportsdata/football/nfl/salaries/team.
6. Players who were not drafted are treated as being drafted after the last pick.
7. It should be noted that individual performance statistics are likely correlated with team per-
formance statistics. In the presence of such multicollinearity, variables which are signifi-
cant may appear insignificant. In order to test for this multicollinearity, models were also
run excluding Win, Winning Percentage, Offensive Points, and Salary Cap Percentage.
Even when excluding those measures from the model individual performance statistics are
found to be insignificant.
References
Arrow, K. J. (1973). The theory of discrimination. In O. Ashenfelter & A. Rees (Eds.),
Discrimination in labor markets (pp. 3–33). Princeton, NJ: Princeton University Press.
Becker, G. S. (1957). The economics of discrimination (2nd ed., 1971). Chicago, IL: Univer-
sity of Chicago Press.
Berri, D. J., & Simmons, R. (2009). Race and the evaluation of signal callers in the National
Football League. Journal of Sports Economics, 10, 23–43.
Conlin, M., & Emerson, P. M. (2006). Discrimination in hiring versus retention and promo-
tion: An empirical analysis of within-firm treatment of players in the NFL. The Journal of
Law, Economics, & Organization, 22, 115–135.
Cox, D. R. (1972). Regression models and life tables (with discussion). Journal of the Royal
Statistical Society, Series B, 34, 187–220.

Gius, M., & Johnson, D. (2000). Race and compensation in professional football. Applied
Economics Letters, 7, 73–75.
Kahn, L. M. (1992). The effects of race on professional football players’ compensation. Indus-
trial and Labor Relations Review, 45, 295–310.
Keefer, Q. A. W. (2013). Compensation discrimination for defensive players applying quan-
tile regression to the National Football League Market for linebackers. Journal of Sports
Economics, 14, 23–44.
Keefer, Q. A. W. (2015). The sunk-cost fallacy in the National Football League: Salary cap
value and playing time. Journal of Sports Economics. doi:10.1177/1527002515574515
Lapchick, R., Donovan, D., Rogers, S., & Johnson, A. (2014). The 2014 racial and gender
report card: National Football League. UCF Institute for Diversity and Ethics in Sport.
Retrieved from http://www.tidesport.org/The%202014%20NFL%20Racial%20and%
20Gender%20Report%20Card.pdf
Mogull, R. G. (1973). Salaries and race: Some empirical evidence. Industrial Relations, 12,
109–112.
Mogull, R. G. (1981). Salary discrimination in professional sports. Atlantic Economic Jour-
nal, 9, 106–110.
Volz, B. D. (2009). Minority status and managerial survival in major league baseball. Journal
of Sports Economics, 10, 522–542.
Author Biography
Brian D. Volz is an assistant professor of economics at Assumption College where he teaches
courses in labor economics and sports economics. He received his PhD in economics from the
University of Connecticut.

Race & Quarterback Survival

Uploaded by

Copyright:

Available Formats

Race & Quarterback Survival

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Race & Quarterback Survival

Uploaded by

Copyright:

Available Formats

Article

Journal of Sports Economics

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Variable and Sample Selection

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

performance must also be included as a covariate. Unfortunately, a widely

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Table 1. Exponential Model.

All Variables Included Significant Variables

Covariate Hazard Ratio P-Value Hazard Ratio P-Value

Black 2.20 .057* 1.42 .041**

proportional hazards is satisfied. The proportional hazards assumption can be tested by

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Table 2. Weibull Model.

All Variables Included Significant Variables

Covariate Hazard Ratio P-Value Hazard Ratio P-Value

Black 2.76 .025** 2.59 .014**

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Table 3. Gompertz Model.

All Variables Included Significant Variables

Covariate Hazard Ratio P-Value Hazard Ratio P-Value

Black 2.56 .032** 1.43 .044**

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Table 4. Cox Proportional Hazards Model.

All Variables Included Significant Variables

Covariate Hazard Ratio P-Value Hazard Ratio P-Value

Black 2.82 .018** 2.88 .005***

1.98 to 2.46 times more likely to be benched than an observationally equivalent

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Exponential Model Weibull Model

Gompertz Model Cox Proportional Hazards Model

Figure 1. Estimated survival functions by race.

Table 5. Predicted Probability of Being Benched.

Cox Model Including All Variables

Week White Black

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Table 6. Difference in PPG Between Replacement and QB.

QBs not Appearing on Injury Report 2001–2009

Mean N Standard deviation

White 1.756 235 9.027

Alternatively quarterbacks may be facing customer discrimination. A major

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

Declaration of Conflicting Interests

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

mikeyhlebasko@gmail.com - November 25, 2018 - Read articles at www.DeepDyve.com

You might also like

Black 2.76 .025 2.59 .014

Black 2.56 .032 1.43 .044

Black 2.82 .018 2.88 .005*