Taking a Knee: Effect of NFL Player Protests on Subsequent Employment and Earnings

Protesters sometimes face penalties for their actions, but few papers have attempted to quantify these penalties. We investigate whether the subsequent salaries and employment status of NFL players who took a knee or sat during the national anthem during the 2017 season differed from similar players who did not. We find limited evidence that they were penalized in terms of employment during the 2018 or 2019 seasons. Conditional on employment, we find an insignificant relationship between protesting and log salaries


Introduction
In 2016, Colin Kaepernick of the San Francisco 49ers protested racism, social injustice, and police brutality against people of color by sitting or kneeling during the performance of the national anthem that typically precedes each NFL game (Hoffman and Minsberg 2018). He has not played professional football since the 2016 season. In 2018, Kaepernick filed a grievance against the National Football League (NFL), accusing the League of colluding to keep him from playing; his case was settled out of court in early 2019.
In a stunning reversal of the NFL's stand on player protests, on June 5, 2020, NFL Commissioner Roger Goodell said, "We, the National Football League, condemn racism and the systematic oppression of Black people. We the National Football League admit we were wrong for not listening to NFL players earlier, and encourage all to speak out and peacefully protest" (NFL.com 2022).
Athletes have a long history of peaceful protests. Protests by athletes are associated with reduced attendance at football games (Watanabe et al. 2019) and lower television ratings (Brown and Sheridan 2020;Miller et al. 2019). Other outcomes affected by athlete protests include social networks (Houghteling and Dantzler 2020), the ethics of patriotism (Montez de Oca and Suh 2020), fan perceptions of athlete activism (Park et al. 2020;Smith and Tryce 2019), White power (King et al. 2007), and market disruption (King and Pearce 2010). Our analysis of the consequences of protesting by NFL players contributes to multiple literatures at the intersection of sports and social sciences, particularly economics.
Although it is difficult to study the economic consequences of a mass protest, El-Mallakh et al. (2018) study the aggregate employment effects of a national-level protest, the 2011 "Arab Spring" protests. They find that the "Arab Spring" Egyptian protests concerning working conditions for women are associated with an increase in Egyptian women's private-sector employment post-protests and a corresponding decrease in Egyptian men's working hours, particularly at lower levels of the pre-revolution income distribution.
In the USA, the summer of 2020 was filled with largescale, organized protests from two different sides of the political spectrum: protests for the Black Lives Matter movement and protests against COVID restrictions. Chenoweth et al. (2022) examine the characteristics of people who attended each type of protest. They document that protestors come from racially and politically diverse backgrounds and "are broadly representative of U.S. citizens on several dimensions" (p. 2). Further, they show that participation in protests appears to be intentional-that "protestors respond to expected costs (e.g., attendance consistent with risk of COVID exposure)" (p. 2). People who attend a protest report views that align with the protests. Taken together, the researchers find evidence that protestors weigh the costs and benefits of protesting and that people anticipate consequences to their participation.
Because of available data on NFL player characteristics, including the names of people who participated in national anthem protests, their salaries, and their work status, we can see if NFL players who participated in public protests suffered any negative labor-related consequences. Like Chenoweth et al. (2022), we study individual protestors, but we focus on the labor-market consequences of protesting at work (on national television) rather than the motivations for protesting.
Our work builds on Niven (2020), who finds that NFL players who protested are more likely to take a pay cut and have lower salary growth. He studied a sample of 200 + players with similar ability as of the 2017 season. In earlier work, Niven (2019) looks at the determinants of NFL players protesting during the 2017 season, not the consequences of protesting. He finds that players with guaranteed salaries and higher-quality players, measured in terms of Pro Bowl appearances, games played in 2017, and being drafted in an earlier round, were more likely to protest. 1 We look at the relationship between NFL players protesting during the 2017 season and their employment status and salaries during the 2018 or 2019 season. Because all the players who protest multiple times are Black, we limit the regression sample to Black players who played during the 2017 season. In most specifications, protesting two or more times during the season is associated with a decline in the likelihood of employment at the start of the 2018 season. This result is consistent with Niven's (2020) analysis of a much smaller sample of players. We find no discernable relationship, however, between protesting and employment at some point during the 2018 or 2019 season; several protestors joined the league during the 2018 season. Similarly, no discernable association exists between protesting and 2018 or 2019 log salary. The pattern of results, for both employment and salary, holds when we include additional controls such as team fixed effects.

Data and Descriptive Statistics
Our focus is on the set of players who played in the NFL during the 2017 season based on data from www. pro-footb allrefer ence. com. These data contain individual player information such as height, weight, and date of birth. These data include draft position for players who entered the league via the draft. For 2018 and 2019, we have data on each player's "salary cap hit," the amount of each player's individual salary that is summed to calculate the team's salary cap.
Data from the Pro Football Reference website also contain player performance information for the 2017 season, including the team, number of games played, number of games started, and a measure of productivity called "approximate value." The approximate value is an integer from 0 to 19, where a higher number represents a more productive player. This statistic is based on both individual performance measures and team performance; it represents an attempt to compare players across positions and across time. 2 For players who have fantasy football statistics (quarterbacks, receivers, tight ends, and running backs), the correlation between approximate value and fantasy statistics for 2017 is over 0.95. In other words, the fantasy football statistic, a strong determinant of player salaries and employment (Jepsen et al. 2021), is highly correlated with approximate value for the 25% of players who have fantasy football statistics. The high correlation supports the assertion that approximate value is a meaningful measure of productivity.
We supplement the 2017 player data with additional data from the Pro Football Reference website. First, we add information on injuries during the 2017 season. Second, we merge the 2017 player data with 2018 and 2019 employment data (by player name and date of birth) in order to identify the subset of 2017 players who are employed in 2018 and 2019. The first employment measure is being on an NFL roster soon after the season began (September 2018). These players have gone through the rigorous pre-season training camp, designed to identify the most promising players for the upcoming season. The second measure of employment is having played in at least one game, and this variable is 1 3 available for both the 2018 and 2019 seasons. 3 For ease of exposition, we will refer to a player who is on the roster at the beginning of the season as employed; similarly, we will refer to a player who has played in at least one game by the end of the season as employed. We are unable to construct a "roster" of players at the end of the season other than by recording those who played in at least one game.
We add data from three additional sources: arrests from USA Today, suspensions from www. spotr ac. com, and race/ethnicity from www. profo otbal llogic. com. We create a dummy variable equal to one for a player who was arrested at any point between the start of his NFL career and July 31, 2018, the beginning of training camp for the 2018 season. The suspension variable is a dummy variable equal to one for players who were suspended during the 2017 season. The researchers at www. profo otbal llogic. com kindly shared data on player ethnicity, which we used to distinguish Black players from non-Black players.
Data on protesting during the national anthem are from the ESPN website. The site maintained a blog during the 2017 season with weekly reports, both in the regular season and preseason, of individuals with any sort of protest during the national anthem. In most instances, the website provides detailed information of each protesting player's actions during the national anthem, such as kneeling versus standing. They also report if the player only protested during part of the anthem. In a few cases, when a large majority of the team protested, the website does not provide individual-level behavior. The appendix lists the instances where individual players are not identified. Most of these instances were in week 3 (September 25), the first game after President Trump repeatedly urged NFL owners to suspend or fire any player who protested during the national anthem.
From the individual player reports, we construct our preferred measure of protesting: a dummy variable equal to one for players who protest more than one time. The focus is on multiple occurrences because the vast majority of players who protested once did so in week 3 following President Trump's statements. Such one-time protests are unlikely to have the same negative consequences compared to the group of players who protest during an average of half the regular-season games. In robustness tests, we consider alternate measures of protesting such as a dummy variable for protesting at least one time or an integer variable for the number of protests. The preferred protest measure is defined as kneeling, sitting, raising a fist, or not being on the field for the anthem. Again, we conduct robustness testing to consider a more inclusive measure of protest that is for protesting or supporting protestors, which includes all the protesting categories as well as players who protested only at the beginning or end of the anthem, who provided physical support for protestors by placing hands on them, or who stood directly next to protesting players as a sign of support (as identified from the ESPN website). Table 1 provides more information on these measures of protest for the entire set of players who were on an NFL roster in 2017. The first two columns are for all players, and the second two are for Black players only. Columns 1 and 3 contain information on protesting, and columns 2 and 4 include displays of support-such as placing a hand on the protesting player's shoulder-as well as protesting.
The top panel demonstrates that 139 players protested at any point during the 2017 season, with only three nonBlack players in that group. Only nine of the 156 players who protested or supported protestors are not Black. The bottom panel of the table illustrates that less than one-third of the players who protested or supported protestors did so multiple times. Only 40 players protested more than one time, and 51 players protested or supported a protestor multiple times. Among players who protested multiple times, the average number of protests  Table 1), we limit the regression sample to Black players. This limitation also allows us to isolate the relationship between protesting and employment from the relationship between race and employment. Table 2 contains descriptive statistics for the regression sample, Black players who played in an NFL game in 2017. The first two columns are for the entire regression sample, the second two columns are for the subset of players who protested multiple times, and the final two columns are for the subset of players who protested once or not at all.
Approximately 58% of Black players in the regression sample were on the same team at the start of the 2018 season and the end of the 2017 season. Nearly 73% were on a roster at the start of the 2018 season, rising to 76.4% at any point during the season before falling to 60% in 2019. Only 3% of players (40 players) participated in multiple protests, but roughly 10% protested at some point during the season.
On average, players have roughly 3 years of experience, and they have an approximate value of 3.2. Roughly 20% of players have an approximate value of zero, and another 20% have an approximate value of one. Only 7% of players have an approximate value of nine or higher. Players played in an average of 11 of the 16 games. Fifteen percent of players were drafted in the first round; over 30% were not drafted. Only 2% were suspended during the 2017 season, and 7% were arrested during their professional career. Defensive backs are the most common position at 26%. One percent of players are either quarterbacks or play on special teams. Because no players protesting multiple *, **, and *** denote statistical significance at the 10%, 5%, and 1% levels, respectively, for a t-test for difference in means between multiple protestors and other players (assuming equal variances) times are quarterbacks or special teams' players, we later test the robustness of our results to the exclusion of these two positions. Players who protest multiple times ("protestors") are more experienced and have more favorable performance characteristics compared to other players, consistent with Niven (2019Niven ( , 2020. Of those on rosters in 2018, they have an average salary of more than three million dollars, compared with under two million dollars for other players. The disparity in mean salary is even larger in 2019. Protestors have an average of more than 4 years of experience compared to three for other players. On average, they have an approximate value of 4.400, in contrast to 3.214 for other players, indicating that their performance is rated higher than the average player in the sample. Protestors also are more likely to be drafted in the first round. They are more likely to have been arrested: 15.0% compared to 6.7% for other players. Despite the generally positive selection on observed performance characteristics, the mean of our dependent variables (playing in the following seasons) is the same for protestors and non-protestors.

Methods and Predicted Effects
The outcome of interest is a measure of employment, either in the 2018 or 2019 season. First, we create a dummy variable equal to one for players who start the 2018 season on the same team as they finished the 2017 season; the variable is zero for players on a different team in 2018 and for players who were not on a roster at the start of the 2018 season. This employment variable looks at the short run to see if players who protested were more likely to leave the team who employed them when they protested. The second dependent variable is a dummy variable equal to one for players who are on an NFL roster at the start of the 2018 season. Although some players have contracts that guarantee a portion of their salary across multiple years, players are not guaranteed employment across seasons (or even weeks within the season). The final two employment measures are dummy variables for playing during the 2018 or 2019 season. Equation (1) depicts the linear probability model we estimate: The key independent variable is a dummy variable equal to one for players who protest multiple times during the season. We test whether the coefficient for this variable, , is less than zero or not with a one-sided test. In other words, we test whether players are less likely to be employed the following (1) Roster = + * Protest + * Productivity + * Characteristics + season if they protest during the national anthem. The results section contains robustness tests using expanded measures of protesting as well as alternate samples. 4 The predicted effect of protesting is negative, given previous research documenting a negative relationship between protesting and player salaries (Niven 2020) and between protesting and lower attendance and television ratings (Watanabe et al. 2019;Brown and Sheridan 2020;Miller et al. 2019).
We attempt to isolate the effect of protesting from other factors by including available measures of productivity and demographics. Approximate value is our primary measure of productivity. It is more appealing than fantasy football statistics or information such as touchdowns or tackles because it is available for every player in the National Football League. We control for experience, measured by the number of years in the league, and for the number of games played during the 2017 regular season. We include dummy variables for players who were drafted in the first round and players who were not drafted, comparing both groups to the omitted category of players drafted in the second through seventh rounds of the draft.
We include two controls for negative behavior. 5 The first is a dummy variable equal to one for players who were suspended during the 2017 season. The second is a dummy variable equal to one for players who were arrested at any point from their entrance into the NFL until the start of training camp for the 2018 season. These two measures are important controls for behavior, especially since, as shown in Table 2, players who protest are more likely to have been arrested. By including these control variables, we do not confound the effects of protesting with the effects of being suspended or arrested.
Even though the dependent variable is binary, we estimate a linear probability model -ordinary least squares models on a dichotomous outcome. Linear probability models have coefficients that are much easier to interpret than logit or probit models, and we show in Table 8 in the Appendix that the results are not sensitive to this choice.
Although the main outcome is employment, we also run a version of the model in Eq. (1) where the outcome is log salary cap hit for 2018 or 2019. Salary cap is the typical measure of salary in sports economics. Note that salary is set to missing for all players not on a 2018 roster, and it is missing for 16 players on 2018 rosters. Because 2019 salary cap is missing for 12% of the regression sample, the results for 2019 log salary should be interpreted with caution. Table 3 contains the results from the linear probability model for the four measures of employment: (1) on the same team at the start of the 2018 season, (2) being on a roster at the start of 2018, (3) playing in a game during the 2018 season, or (4) playing in a game during the 2019 season. In addition to the variable of interest, protesting, the regression also includes controls for player productivity, including measures of suspensions, arrests, and individual positions such as linebacker or running back.

Results
Although we cannot discern an association between protesting multiple times and being on the same team at the start of the 2018 season, we find that a player who protests multiple times has a lower likelihood of being on a roster at the start of the following season of 10.9 percentage points. This association is significant at the five-percent level for a one-sided test. Taken together, these two results are consistent with stronger league-level negative consequences of protesting compared with team-level.
As in previous work (Jepsen et al. 2021), we find a positive relationship between productivity and employment. The coefficient for approximate value is positive; the coefficient for approximate value squared is negative, suggesting diminishing returns to productivity. A one-standard deviation increase in approximate value evaluated at the mean value corresponds with a 14 to 17 percentage-point increase in the likelihood of employment. Each year of experience reduces employment in 2018 (columns 2 and 3) by two percentage points, an amount equal to the increase in employment associated with playing in another game. Being a first-round draft pick is associated with an increased probability of employment of 7 to 11 percentage points, whereas not being drafted is associated with a decreased probability of employment of 4 to 8 percentage points. Hence, the coefficient for protesting in column 2, − 0.109, is large relative to the other coefficients in Table 3. Table 8 in the Appendix illustrates that the results for this outcome are similar across logit, probit, and linear probability models.
However, protesting does not have a discernable effect on the likelihood of playing at any point during the 2018 or 2019 season. As shown in columns 3 and 4, the coefficients on protesting are now only − 0.033 for 2018 and − 0.032 in 2019 and are not statistically significant at 10% in a one-sided test. Thus, Table 3 Linear probability models for employment in 2018 or 2019 season Each column is from a separate OLS regression. Robust standard errors are in parentheses. *, **, and *** denote statistical significance at the 10%, 5%, and 1%, respectively, for a two-sided test; + and + + denote significance at the 10% and 5% levels, respectively, for a one-sided test for the coefficient on protesting being negative. In addition to the variables shown, the results include dummy variables for each of the following positions relative to the omitted position of wide receiver: offensive lineman, tight end, running back, quarterback, defensive lineman, linebacker, and defensive back the coefficient on protesting declines by close to eight percentage points (or 70%) compared with the outcome of being on a roster at the start of the 2018 season. The pattern of results in Table 3 is consistent with NFL teams preferring players who have not protested over those who have when choosing their teams at the start of the 2018 season. Once teams are scrambling to find players to replace injured players during the season, however, they appear to be much less concerned with a player's protest history. Alternatively, perhaps teams were less concerned about the players' social statements once the 2018 season started because almost none of them protested in 2018.
The results in Table 3 use protesting multiple times as the measure of protest. In Table 4, we explore alternate definitions of protesting, considering six different measures of protesting: (1) multiple protesting, the preferred measure as in Table 3; (2) any protesting; (3) the number of protests; (4) multiple protesting or supporting; (5) any protesting or supporting; and (6) the number of games protesting or supporting. Thus, rows 1, 2, and 3 focus solely on protestors and assume that supporting has no effect on employment. Rows 4, 5, and 6 combine protestors with players supporting the protestors. The columns contain the results for the four employment outcomes, as in Table 3. Each coefficient and standard error are from a separate regression.
In columns 1 and 2, the short-run employment outcomes, the effect of protesting is sensitive to how we classify protesting. The results are most pronounced for players who protest multiple times. In row 2, the variable for protesting at least one time is small in magnitude (around 0.02) and is not statistically significant from zero. When we measure protesting as the number of weeks protesting (row 3), the effect of protesting one week is − 0.007 and is not statistically different from zero at the 10% level for a one-sided test.
The results in Table 4 are robust to whether we broaden our measure of protesting to include players who support protestors.
When we consider being on a roster at the start of 2018 (column 2), the coefficient from the preferred model in Table 3 is − 0.109, compared with − 0.098 when protesting is measured as protesting or supporting protestors multiple times. The coefficients from these two specifications are similar for other outcomes as well. None of the other measures of protesting-rows 2, 3, 5, and 6-is statistically significantally different from zero at 10% (for a one-sided test), and the largest coefficient is only 4.3 percentage points in magnitude (row 2, column 4).
In sum, the results in Table 4 suggest that there are adverse employment consequences for players who repeatedly protested during the national anthem in 2017 but only for being on a roster at the start of the 2018 season. When the only control for protesting is a dummy variable for any player who protested, regardless of how many times, the coefficient is statistically and economically insignificant.
Our final alternative measure of protesting is to estimate a model with separate coefficients for protesting one time and for protesting multiple times. As shown in Table 9 in the Appendix, the coefficient for protesting one time is positive in all eight specifications, clearly showing that players who protested once did not suffer adverse employment outcomes. The coefficient for protesting multiple times is only slightly smaller in magnitude once we add a control for protesting once. Furthermore, we can reject the hypothesis that the effects of protesting once are the same as protesting multiple times, providing support for the decision to treat single protestors and multiple protestors differently.
A potential concern with the results is that they are specific to the set of variables or observations used. To address this concern, Table 5 includes results from several different specifications. For example, we test whether the results are Each coefficient and standard error are from a separate OLS regression, for a total of 24 regressions. Robust standard errors are in parentheses. + and + + denote significance at the 10% and 5% levels, respectively, for a one-sided test (of the coefficient being negative). All specifications include controls for productivity and demographics as shown in Table 3 Table 3. The table illustrates the robustness of the findings: a significant decline of roughly 11 percentage points in employment at the start of the 2018 season and an insignificant decline of roughly 3 percentage points in employment at any time during the 2018 season. In other words, the decline in employment at the start of the season associated with protesting multiple times cannot be explained by having too many controls, such as for position. Nor can the results be explained by excluding additional player characteristics (such as being foreign born or injured) or team characteristics (such as having a Black head coach, having a coaching change, or team performance). The only insignificant coefficient in the left panel is for the specification with all the 2016 variables on the reduced sample of players on 2016 as well as 2017 rosters.
Of particular interest is the similarity of the results to the inclusion of team fixed effects. In other words, the coefficient is similar whether we compare multiple protestors to their teammates (e.g., the within-team variation in protesting) or to players on any team (e.g., the within-team and the betweenteam variation in protesting).
Another notable finding is that the results are qualitatively similar if we use 2016 measures of a player's performance rather than 2017 measures in order to avoid potential bias if 2017 performance is affected by protesting. Because we limit the sample to players on a 2016 NFL roster, however, the estimates become less precise as a consequence of reducing the sample size from 1403 to 1041.
For the second outcome, playing at any point during the 2018 season, none of the coefficients is statistically significant at the ten-percent level for a one-sided test. Most of the coefficients are around − 0.03. However, the results nearly Table 5 Alternate specifications for employment during 2018 season Each row and panel contain a separate OLS regression. + and + + denote significance at the 10% and 5% levels for a one-sided test (of the coefficient being negative). All specifications include controls for productivity and demographics as shown in Table 3. Unless otherwise stated, the number of observations in each regression is 1403 a These regressions exclude quarterbacks and special teams players, so the number of observations is 1389 b These regressions exclude players who did not play during the 2016 season, so the number of observations in each regression is 1041 c These regressions do not include controls for position due to the small number of players on a given team double in magnitude, to − 0.060, when the model contains team fixed effects. The coefficients are − 0.080 and − 0.073 for models using 2016 player characteristics, where the sample is restricted to players who played in the 2016 season. Even though we cannot reject the hypothesis that protesting multiple times in 2017 is unrelated to playing during the 2018 season, the size of the coefficient for multiple protesting depends on the control variables included to a much greater extent than for the outcome of being on a roster at the start of the 2018 season. Table 10 in the Appendix contains the results for two additional employment measures: being on the same team at the start of the 2018 season and playing during the 2019 season. For being on the same team, most coefficients are non-trivial in magnitude, but only two out of 23 are statistically significant at the 10-percent level for a one-sided test. Thus, the relationship between protesting and switching teams before the start of the 2018 season is inconclusive. In contrast, the results for playing during the 2019 are consistent across all specifications: no detectable relationship between protesting and being on a roster in 2019. The coefficient is between − 0.019 and − 0.045, with standard errors of at least 0.064.
To address the concern that players do not protest at random, we conduct matching estimation using five different estimators. We compare players who protest multiple times with the subset of players who do not protest multiple times. The first two techniques use the likelihood of protesting, calculated using a logit model, to match multiple protestors with other players. The first technique calculates inverse probability weights based on this likelihood, whereas the second technique, the propensity score method, uses nearest-neighbor matching to select a comparison group member (a player who did not protest multiple times) with the most similar likelihood of protesting. Both methods use a common support restriction. 6 The final three matching techniques use Mahalanobis matching, a technique that calculates the distance between players based on covariates (without any regard for the likelihood of protesting). We include two versions of Mahalanobis matching where we do exact matching on quartiles of experience or games played. 7 For propensity score Malahanobis matching, we match with replacement, meaning that a player who did not protest multiple times may be matched with multiple protestors. 8 Table 6 contains differences in means between multiple protestors and the matched comparison group of players who did not protest multiple times. In addition, the top row of the table contains the OLS results from Table 3 for comparison purposes. For the outcome of starting the 2018 season (column 2), all but one of the matching estimators are small and statistically insignificant, casting doubt on the robustness of the negative and significant coefficient from the OLS regression. For the remaining outcomes, none of the results, either for OLS or for matching, is negative and statistically significant, which is consistent with a failure to find evidence of any negative consequences of protesting. For the matching estimators, standard deviations are in parentheses, and they are calculated using the Stata "teffects" command. See Table 3 for more information on the OLS coefficients and estimation. + and + + denote significance at the 10% and 5% levels, respectively, for a one-sided test (of the coefficient/mean being negative) So far, the focus has been on employment. In Table 7, the dependent variable is the natural log of the player's salary cap value; the left columns are for 2018, and the right columns are for 2019. For players not on a roster at any point during the relevant season, the salary cap value is set to missing. 9 Salary cap information is also missing for 16 players who played in 2018 and for 105 players (roughly 12%) who played in 2019. Table 7 follows the same format as Table 5 except for the change in the dependent variable. Table 7 shows that the variable for protesting multiple times has no discernable relationship with log salary cap. For 2018, the coefficients range from − 0.013 to 0.048, with standard errors around 0.15. Thus, the coefficients are never statistically different from 0 at the 10% level for a one-sided test. For 2019, all the coefficients are positive and are at least 0.096 log points, consistent with no adverse salary consequences of protesting.

Conclusion
This paper provides new evidence of the effects of protesting on protestors' individual labor-market outcomes. Specifically, we study the relationship for professional football players using the drastic increase in protesting during the 2017 season. In our preferred OLS model, protesting multiple times during the national anthem in 2017 is associated with a decline in the likelihood of being employed at the start of the 2018 regular season, but the Each row and panel contain a separate OLS regression. + and + + denote significance at the 10% and 5% levels for a one-sided test (of the coefficient being negative). All specifications include controls for productivity and demographics as shown in Table 3  result is usually insignificant when we use matching estimators rather than OLS. Protesting multiple times is associated with a much smaller and statistically insignificant decline in playing at any time during the 2018 or the 2019 season. Similarly, we find no evidence of an association between protesting and log salaries. The pattern of results is robust to alternate definitions of protesting, as well as to the inclusion of additional control variables such as team fixed effects. Protesting one time, which many players did in week 3 in response to President Trump's multiple statements against protests, is not associated with a decline in any of our measures of employment. One caveat is that we treat protesting as an exogenous variable, as we are unable to control for the potential endogeneity of protesting beyond the inclusion of additional control variables (either in OLS models or matching estimators).
The results suggest that protesting injustices by high-profile individuals can result in adverse employment outcomes in the short run, but the long-run evidence is less conclusive. Future work should look at longer-run outcomes for these individuals. Given the recent global protests and the NFL's admission that it failed to listen to its Black players, the upcoming seasons should provide an important opportunity to compare future protests to the 2017 season to see if negative employment effects disappear. In general, more research is needed on the labor-market consequences of protesting, especially in other contexts beyond professional sports.

Appendix. List of games with missing protest information
The following games do not have information on the specific players who protested the national anthem. Thus, all variables measuring protest behavior do not include protests from these games.
Week 5 (October 1, 2017) San Francisco 49ers. Each column is from a separate model. The logit and probit models report marginal effects. Robust standard errors are in parentheses. + and + + denote significance at the 10% and 5% levels, respectively, for a one-sided test (of the coefficient being negative); *, **, and *** denote statistical significance at the 10%, 5%, and 1%, respectively, for a two-sided test. All specifications include controls for productivity and demographics as shown in Table 3 Logit Probit OLS  Each column is from a separate OLS regression. Robust standard errors are in parentheses. + and + + denote significance at the 10% and 5% levels, respectively, for a one-sided test (of the coefficient being negative). All specifications include controls for productivity and demographics as shown in Table 3 2018 Each row contains a separate OLS regression. + and + + denote significance at the 10% and 5% levels for a one-sided test (of the coefficient being negative). All specifications include controls for productivity and demographics as shown in Table 3. Unless otherwise stated, the number of observations in each regression is 1403 a These regressions exclude quarterbacks and special teams players, so the number of observations is 1389 b These regressions exclude players who did not play during the 2016 season, so the number of observations in each regression is 1041 c These regressions do not include controls for position due to the small number of players at each position on a given team