Reply to:
Differences in the Success of NFL Coaches by Race:
A Different Perspective
Janice Fanning Madden
Matthew Ruther
University of Pennsylvania*
*We thank the Wharton Sports Business Initiative (WSBI) for support for this project.
0
Abstract
Malone, Couch and Barrett (2008) argue that a broader analysis of fired coaches, including
adding in partial season fires, considering a wider range of causes of firing, and analyzing rehiring makes
the results reported by Madden (2004) “disappear.” We show that Malone et. al. have analyzed
inaccurate data. When the data used by Malone et. al. are corrected and their speculations tested
empirically, Madden’s conclusion that analyses of all employment decisions involving head coaches
between 1990 and 2002 are consistent with discrimination against African Americans is supported.
Madden (2004) analyzed several dimensions of NFL coaching records for the 1990-2002 seasons
(prior to the 2003 adoption of the Rooney rule which required teams to interview minority candidates
before appointing head coaches) and established that African American coaches performed better than
their white counterparts. She interprets this evidence as consistent with the hypothesis that African
American coaches were being held to higher standards to get their jobs in the NFL. Malone et. al.
(2008) argue that adding partial seasons to the analysis of one dimension of NFL coaching records,
performance record when fired, changes the result for the analysis of firing. Malone et. al. also examine
the rehiring of fired coaches and propose other explanations for the low representation of African
Americans among NFL head coaches. We respond to Malone et. al. below.
Analysis of Fires
Madden (2004) presents evidence that, between 1990 and 2002, African American full season
NFL coaches were significantly more likely to be fired given their season records than their white
counterparts. Malone et. al. (2008) argue that the evidence changes if coaches fired mid-season are
added to the analysis. They also imply that an analysis with partial season coaches is preferred. Table 1
lists the partial season coaches and their records -- that is, the 11 teams and years (accounting for 22
1
coaches) that had two coaches in one season between 1990 and 2002.1 Malone et. al. base their
conclusion that partial seasons change the results on a regression analysis that adds only three “net”
coaches to the Madden (2004) analysis. There are 374 coaches with full season records2 and Malone et.
al.’s regression analysis includes only 378 observations. These observation counts show that Malone et.
al. have not analyzed an accurate data set.
Table 2 reports estimates of six race effects (both coefficients and z scores) for two different
specifications of the firing logit regression, each estimated using three different samples of coaches.
The first result, full season coaches with season record measured by total wins (first row and column)
shows the Madden (2004) result that African American coaches are more likely to be fired. The next
two rows show the effects on the racial coefficient of including partial season coaches. First, we add the
9 partial season head coaches who start the season and are fired mid-season3 (and not the interim
coaches who finish the season) to the coaches included in the first row analysis. Then, we add all partial
season coaches, including the interim coaches hired only to finish the season, to the coaches included in
the second row analysis. Because interim coaches are hired only to complete the season, they cannot
be fired from a regular coaching job. Furthermore, these interim coaches inherit troubled teams that
are likely to perform poorly. Including interim coaches as candidates for firing who effectively cannot
be fired and who have poor performance records is questionable at best. When head coaches under
1
Malone et. al. (2008) correctly point out that Madden included the 2001 Minnesota Vikings season as a
full season with one coach, Dennis Greene, even though he was fired right before the Vikings lost their final season
game. The exclusion of Dennis Greene’s firing would reduce the number of African American fired coaches from 5
to 4. No other partial season coach had 15 games. As the final game was a loss, attributing that game to Dennis
Greene does not improve his record at fire. Madden included Dennis Greene as a full season coach, then, because
African American observations are needed for a statistical analysis of race to have statistical power and the
inclusion of Mr. Greene does not introduce any bias toward finding African Americans were performing better than
whites at fire.
2
The Madden 2004 analysis had 375 full season coaches because Dave Shula was mistakenly included as a
full year coach for the Cincinnati Bengals in 1996; there were also some minor corrections to win-loss records.
3
Of these 11 white coaches, nine depart involuntarily (i.e., are fired). Bruce Coslet and Bobby Ross resigned
mid-season.
2
contract who depart mid-season are included in the analysis (second row of the first column), there is
very little change in the result: African American coaches are still more likely to be fired and the
difference is significant at a 0.95 probability standard. When interim coaches are added, the race
coefficient drops, but is still positive and the z score drops from 2.02 to 1.46. Because there are only
thirty total African American coach-year observations -- five of whom are fired -- in these analyses, it is
surprising that race is statistically significant with any specification. A move of 0.56 standard deviations
with these small numbers of African Americans is hardly surprising.
There are several reasons why an analysis that includes mid-season fires is not preferred over an
analysis of coaches fired at the end-of-season. The current season records for coaches fired midseason include anywhere from 5 to 13 games and for interim coaches the number of games range from
3 to 13. Madden (2004) and Malone et. al. (2008) measure current season performance using the
number of wins and whether the team made the playoffs. Coaches fired mid-season cannot have made
the playoffs and their win counts are not comparable indicators of performance relative to full season
coaches. A better way to measure current season performance when partial seasons are included is to
use win percentage, rather than the number of wins. The second column of Table 2 repeats the
analyses of the first column but replaces the number of wins with the percentage of current season
games won. The third row, where 22 partial season coaches are added to the analysis, has a noticeable
change from the first column results. The coefficient and z statistic for race of 0.761 and 1.84 are much
more similar to the results when fewer or no partial season coaches are included in the analyses.
Other issues remain, however, when partial season coaches are included in the analysis of fires.
If mid-season fires occur under different circumstances than end-of-season fires, then their
determinants may be systematically different. For example, end-of-season fires may be less likely than
mid-season fires to reflect an individual game outcome. Other issues arise when interim coaches are
added to the mix. These coaches have even fewer games, and fewer wins, but cannot be fired.
3
Furthermore, between 1990 and 2002, only nine (out of 68) fires of NFL coaches occurred in mid-season
and all were white. Introducing partial season coaches, especially interim coaches, clearly obfuscates
the analysis of the role of race in firing
Malone et. al. identify a variety of circumstances that may lead a team to fire the head coach,
including the cumulative record or a “plateauing-out”effect.4 The last columns of Table 2 repeat the
fire analyses but add controls for cumulative records and a plateauing-out effect. We add cumulative
win percentage up to current season for each coach, cumulative proportion of seasons in the playoffs,
and controls for the differences between the cumulative team records and the current season record.
In each case, the racial coefficients increase, although the z statistics decline due to the decreasing
degrees of freedom.
Of course, Malone et. al. are correct that the decisions of NFL teams (as well as the economic
decisions of other firms, consumers, employers, etc.) involve a variety of concerns and tradeoffs. Such
complexities have not stopped economists from estimating wage equations, hedonic equations for
houses or art, etc. The constraints on analysis implied by the Malone et. al. discussion in the case of NFL
coaches, however, are reminiscent of those critics of empirical social science who assert that since each
outcome is unique (or has unique components), there can be no quantitative analysis of outcomes.
Surely the fact that performance records and experience predict which coaches are fired undercuts the
suggestion that fires cannot be analyzed statistically.
Analysis of Rehires
Malone et. al. assert that all of the African American head coaches were rehired after being
fired. They then raise two issues about the rehiring of NFL coaches who were fired, in the context of
4
We note, however, that Fee et. al. (2006) report in their study of NFL head coach firings between 1970
and 2001 that the team’s win-loss record is the most frequently mentioned reason. They report two cases where
playoff success is the deciding factor. They note that lack of respect from players in frequently cited. They do not
report any evidence for the other reasons cited by Malone et. al.
4
Madden’s study of racial disparities in coaching performance. First, Malone et. al. argue that the
rehiring of African American coaches shows that the low representation of African Americans among
head coaches is due to “inertia” and not to racial discrimination. Second, rehiring itself is an
employment decision that was not analyzed in Madden (2004).
The statement that all African American coaches who were fired were rehired is not quite
correct. Of the five African Americans who were fired, Art Shell was not rehired as a head coach again
for 12 years after he was fired from Oakland in 1994 and Ray Rhodes has not been rehired after being
fired from Green Bay in 1999. Malone et. al. also assert that the rehiring of African American head
coaches shows that it was lack of experience rather than race that accounts for the underrepresentation
of African Americans among head coaches. Yet, Malone et. al. perform no analyses to show that prior
experience differentials by race among coach candidates account for the stronger performance of
African American coaches, which Madden has interpreted as being consistent with African Americans
being held to higher standards to be hired as coaches. In fact, the performance advantages of African
American coaches demonstrated in Madden (2004) also occur if the comparisons are restricted to
coaches in their first coaching jobs (that is, eliminating all rehires, whatever the reason for leaving the
prior coaching job). Table 3 shows the difference in performance by race in the first NFL coaching job.
African American coaches perform better than white coaches, as measured by wins and proportion of
seasons in playoffs, among coaches with no prior NFL head coach experience. Regression analyses
which add controls for years with team and player salaries also show significantly better regular season
win and playoff records for African American NFL head coaches with no prior head coaching experience.
These results are comparable to those shown for all coaches in Madden (2004).
Rehire is, however, an aspect of head coaching that is relevant to an analysis of racial disparities.
African American coaches who were fired were more likely to be rehired (60% versus 22.2%) within
three years of being fired. African American coaches may have been more likely to be rehired because
5
they had stronger records when they were fired. The three African American rehires had won 54.4% of
their total games with their previous team and went to the playoffs 65.6% of the seasons; the 14 white
rehires had won insignificantly fewer games (51.9%, 0.57 standard deviations), but participated
significantly less in the playoffs (34.6%, 2.35 standard deviations). A logit regression of rehire on race,
win percentage, and years as NFL head coach for all coaches fired between 1990 and 2002 showed a
positive insignificant, or favorable (1.18 standard deviations), effect of being African American.
Malone et. al. state that the rehire of African American coaches who are fired “belies” racial
discrimination. We don’t understand their logic and they give no further explanation. We believe the
facts to be consistent with implicit racial discrimination in hiring.5 NFL teams are not consciously
rejecting African American coaches, but are unconsciously or implicitly discounting them in the presence
of uncertainty and ambiguity as to who will be a successful coach for the initial hire as an NFL head
coach. When the evidence of quality of performance is clearer (that is, less ambiguous or more
explicit), African Americans are rehired as head coaches; when the evidence is ambiguous, as in the case
of initial hire, rather large racial disparities appear.
Conclusions
African American head coaches are more likely to be fired, given their current season records.
Analyses controlling for cumulative records show similar results. African American coaches who are
fired are more likely to be rehired, although there is evidence that the rehired African American coaches
have equivalent season win and stronger prior playoff records than rehired white coaches. The
evidence on fire and rehire, when combined with the other evidence of greater successes for African
5
There is a vast literature in psychology, sociology, law, and more recently in economics on unconscious
bias or implicit discrimination. See Bertrand et. al. (2005) for an overview and Price and Wolfers (2007) for a
recent application to the analysis of racial disparities in call by referees within the National Basketball Association.
6
American coaches presented in Madden (2004), is consistent with implicit racial discrimination in the
NFL during the 1990-2002 period.
References
Bertrand, M., Chugh D., & Mullainathan, S. (2005) “Implicit Discrimination,” American Economic Review
Vol. 95, No. 2, pp. 94-98.
Fee, C.E., Hadlock, C.J., & Pierce, J.R. (2006) “Promotions in the Internal and External Labor Market:
Evidence from Professional Football Coaching Careers,” Journal of Business, Vol. 79, No. 2, PP. 821-850.
Madden, J. F. (2004) “Differences in the Success of NFL Coaches by Race, 1990-2002,” Journal of Sports
Economics, Vol. 5, No. 1, pp. 6-19.
Malone, K.D., Couch, J.F., & Barrett, J.D. (2008) “Differences in the Success of NFL Coaches by Race: A
Different Perspective,” Journal of Sports Economics, Vol. 9, No. 6, pp. 663-670.
Price, J. & Wolfers, J. (2007) “Racial Discrimination among NBA Referees,” (Cambridge, MA: National
Bureau of Economic Research Working Paper 13206.
7
Table 1
List of Partial Season Coaches, 1990-2002
(First coach listed for each team was fired (or resigned); subsequent coach was interim)
Year
1990
Team
Cleveland Browns
1990
Tampa Bay Buccaneers
1991
Indianapolis Colts
1994
Tennessee Titans
1996
New Orleans Saints
1996
Cincinnati Bengals
1998
San Diego Chargers
2000
Arizona Cardinals
2000
Cincinnati Bengals
2000
Detroit Lions
2000
Washington Redskins
Coach
Bud Carson
Jim Shofner
Ray Perkins
Richard Williamson
Ron Meyer
Rick Venturi
Jack Pardee
Jeff Fisher
Jim Mora
Rick Venturi
Dave Shula
Bruce Coslet
Kevin Gilbride
June Jones
Vince Tobin
Dave McGinnis
Bruce Coslet*
Dick LeBeau
Bobby Ross*
Gary Moeller
Norv Turner
Terry Robiskie
*Resigned, was not fired
8
Wins
2
1
5
1
0
1
1
1
2
1
1
7
2
3
2
1
0
4
5
4
7
1
Losses
7
6
8
2
5
10
9
5
6
7
6
2
4
7
5
8
3
9
4
3
6
2
% Wins
0.22
0.14
0.38
0.33
0.00
0.09
0.10
0.17
0.25
0.13
0.14
0.78
0.33
0.30
0.29
0.11
0.00
0.31
0.56
0.57
0.54
0.33
Table 2
Determinants of Being Fired, NFL Coaches, 1990-2002
Controlling* for Record with
Season Total Wins
Season Percentage
Wins
Cumulative Win
Percentage and
Change in Current
Season
Number of
Coefficient
Coefficient
Coefficient
Seasons
coachfor Coach
for Coach
for Coach
Included in
season
is African
z
is African
z
is African
z
Analysis
observations American statistic American statistic
American statistic
Full season
coaches only
374
0.897
2.17
0.910
2.15
1.114
1.87
All, except
interim coaches
383
0.869
2.02
0.838
1.95
1.049
1.77
All coaches
396
0.654
1.46
0.761
1.84
1.69
0.937
*All regressions also include the following independent variables: whether in playoffs, total years with team,
and the square of total years with team. z statistics are computed with robust standard efforts, clustering
on individual coaches.
9
Table 3
Average Season Wins and Likelihood of Playoffs
for Full Season NFL Coaches in Their First NFL Head Coaching Jobs by Race, 1990-2002
Average wins per year
African-American Coaches
(26 full season observations)
9.3
White Coaches
(170 full season
observations)
7.8
t statistic for
difference by race
2.63
Proportion in playoffs
73%
35%
3.69
10