library(tidyverse)
#load the final dataset
Total_Data <- read_csv(file='https://raw.githubusercontent.com/beatricemagistro/Racial_Bias_SerieA/main/Final_Merged.csv')
# create some new variables
Total_Data$COVID <- ifelse(Total_Data$Season.x=="2020-21", 1,0) # dummy for COVID
Total_Data_subset<- subset(Total_Data, Position != "GK") #exclude GKs because practically no fouls and no tackles
Total_Data_subset$Position <- as.factor(Total_Data_subset$Position) #transform to factor
Total_Data_subset$Season.x <- as.factor(Total_Data_subset$Season.x) #transform to factor
Total_Data_subset$Squad.x <- as.factor(Total_Data_subset$Squad.x) #transform to factor
# additional variables (DT)
Total_Data_subset$it = ifelse(Total_Data_subset$Nation=='ITA', 1, 0) #dummy for Italian nationality
Total_Data_subset$Italian= as.factor(Total_Data_subset$it) #make it a factor
Total_Data_subset$cards = Total_Data_subset$yellow_cards_overall + 2 * Total_Data_subset$red_cards_overall #total number of cards, with reds counting doubleRacial Bias in Italian Football Revisited
Introduction
A recent article in the journal Sociology claims that ‘darker-skinned [football] players receive more foul calls and more cards than lighter-skinned players’. The authors - Beatrice Magistro and Morgan Wack - collect data on 2205 football players in the Italian Serie A for the period 2009-2021. The dataset includes information on the skin color of the player (from the Football Manager videogame), number of attempted tackles, fouls committed, yellow and red cards received, minutes played, position, nationality, and club. Analyzing this data, the authors find ‘evidence of systematic racial bias against dark-skinned players’. They conclude that ‘the effects of racial biases in the Italian Serie A are striking’ and ‘[D]arker-skinned players are judged to have been guilty of both minor and major infractions at higher rates than their lighter-skinned teammates’. These are clearly important results on a normatively and politically salient topic, so it is not surprising that the article received coverage in the popular press (e.g. in The Guardian, see also the Twitter thread).
I am always interested in how we can establish bias and discrimination. The methodological challenges involved are inevitably considerable, and this article presents a carefully-argued and executed approach. The authors conduct a number of robustness checks, make many sensible statistical modeling choices, and openly discuss the limitations and uncertainties of their work. Best of all, they have made their data and analysis scripts public and openly accessible. Kudos for that! With the data being available, I decided to take a deep dive and see what patterns I find for myself.
Overall, my conclusions are rather different that the ones advanced in the article. I find that there are no significant associations between skin color and fouls/cards, when we only include appropriate controls. In fact, the number of fouls/cards vary in curvilinear fashion with skin color, so that players in the middle of the skin color distribution received the most fouls and cards. When potentially bad controls are included, for the players with Italian nationality, the number of fouls/cards increases significantly with skin color, but the effect is driven almost entirely by a couple of players - namely, Mario Balotelli and Stefano Okaka. Once we remove these two players from the data, there is no significant effect of skin color on fouls/cards for the Italian players as well.
In sum, it is very hard to interpret the evidence in racial terms. If it is about race, how come darker skin is associated with more negative referee decisions only for the Italian nationals (if we keep Balloteli and Okaka in the sample and conditions on variables such as minutes played, which - as I explain below - are possible colliders), but not for all the foreigners in the league? Why is the link between skin color and referee decisions more likely to be found for players with the same nationality than for players with different nationalities, given that race (whatever it means) varies much more strongly between than within nationalities? Moreover, given that the number of fouls/cards peak in the middle of the skin color distribution, what race corresponds to that?
Finally, I see no evidence that the effects of skin color on fouls/cards were significantly different during the 2020/21 season, which was played at empty stadiums due to COVID-19. This goes against the claim of the article that the public is responsible for a significant part of the racial bias in referee decisions.
Does this mean that there is no racial bias in the Italian Serie A? Of course not! Absence of evidence is not the same as evidence of absence. But - in my interpretation - the available data does not justify the strong claims about racial bias advanced in the article.
I present my analysis below. Comments and disagreements are welcome. I would like to thank again the authors for making their data and code available, which makes this re-evaluation of their arguments possible.
Data Analysis I: Descriptive analysis
To start with, let’s load and prepare the original dataset (following the script of Magistro and Wack, and adding a couple of additional variables).
There are 5990 observations in the dataset (after removing the goalkeepers), but they correspond to 2034 unique players, because the unit of observation is a player-during-a-season.The skin color variable ranges between 1 and 20, with a median of 5 and a standard deviation of 4.4. The interquartile range is of length 4, between 4 and 8. Here are photos of some of the players with skin color values of 1, 5, 10/11 and 19/20.
Total_Data_subset$Player.x[which(Total_Data_subset$Skin==min(Total_Data_subset$Skin, na.rm=T))]
Total_Data_subset$Player.x[which(Total_Data_subset$Skin==median(Total_Data_subset$Skin, na.rm=T))]
Total_Data_subset$Player.x[which(Total_Data_subset$Skin==10)]
Total_Data_subset$Player.x[which(Total_Data_subset$Skin==max(Total_Data_subset$Skin, na.rm=T))]
