Ethnic Discrimination in the Labor Force
STAT 272 Final Paper
Madeline Chang, Elizabeth Murphy, Tra Pham
Introduction
For centuries, ethnic discrimination has been a major issue in many countries. In society today, discrimination in the labor market leaves many qualified citizens unemployed. However in the past it has been difficult to prove these inequalities are the result of discrimination. In this research project, we aim to communicate which characteristics on resumes are statistically significant in showing prejudice in the Boston and Chicago area when applying to jobs.
Relevant literature in this area of investigation includes evidence of ethnic discrimination. In a study done in 2007 in the Stockholm area, resumes for pairs of equally merited applicants were submitted to job openings. By comparing job interview offers for applications with Swedish sounding names and foreign sounding names. By measuring callback rates, it was concluded that applications with Swedish sounding names were chosen over Arabic or African sounding names in most occupations studied. It was determined that only two out of fifteen occupations did not have data that was statistically significant to conclude name discrimination. [2] Another piece for relevant literature focused on how implied characteristics that follow from names often affect who is hired for a job. In 2003, Bertrand and Mullainathan examined the effect a name can have on call back rates. This study found the greater quality resume did not affect who received a call back when paired with an African-American name. [3]
Since our research question is The question that we will investigate with our analysis is: does the race assigned to a name on a resume and callback rate have a correlation? Using logistic regression and other variables to improve our race model, we aim to discover if these predictor variables are significant in callback rates. We hypothesize that race and callback are correlated and that some of the predictor variables in the model will be significant in callback rates.
Materials and Methods
The data we are using to test our hypothesis is from a previous study conducted in 2004 that focuses on ethnic discrimination in the labor market. In this study, fabricated resumes are sent out to help-wanted ads in the newspapers of these cities. Responding to over 1,300 employment ads, nearly 5,000 invented resumes were sent out to these ads. By manipulating perceived race, resumes are randomly assigned African-American or White sounding names. The other predictors of the study are also randomly assigned to each resume, independent of any race assignment. To reinforce the quality gap between two equally merited resumes, labor market experience, career profile and skills listed are added subtly to each resume. For each ad, the study sent in four resumes that matched the job description, two high quality and two low quality. For each quality level of resume, African-American or White names were randomly assigned. The study recorded which resumes received a call back from the ad as well as the corresponding variables that were assigned to the application. [1]
Along with randomly assigning perceived race, other variables we will be using from the original study in our exploration include sex and education. We also will be using whether the ad mentions some educational requirement or “equal opportunity employer”. Our measured categorical binary response variable is call, “Yes” received a callback while “No” did not receive a callback.
The experimental sampling design consists of grouping subject by race and assigning sex, education level, educational requirement, and ‘eoe’. There is no control group in the data. When collecting the data from the original dataset, there are many variables we did not use due to missing data and the lack of need for them in our analysis. Because of this, we did not have to modify our data to account for missing values and our sample size is the same: 2435 resumes for white names and 2435 resumes for African-American names. Our total sample size is 4870 resumes.
Since our research question is does the race assigned to a name on a resume and callback rate have a correlation, we will first look at important summary statistics for race in callback rates. Data will be put in 2x2 tables to find the percentage of each variable that received a callback to determine the relationship between callback and race. Since the response variable is categorical, it will also be fit to logistic regression models to determine significance of variables in callback rates. Various models will be fit to assess the significance of the coefficients. An ROC curve with an AUC score will be fit to the best-fitting model to assess its ability to accurately predict the data within this dataset.
Results
Figure 1: Segmented bar plot to compare percentages within race
Figure 1 is a segmented bar plot that helps compare percentages of each race that received a callback. This figure gives us a good initial look into our hypothesis. Taking into account that each data value is a 0 or 1 for race, the mean value for an Africa-American name is 0.0645 while the mean value for a white name is 0.0965. The bar plot also supports these statistics, we see that the percentage of applicants that received a callback for white names is bigger than the percentage for African-American names.
Table 1: Mean callback rates for all resumes
Table 1 summarizes average callback rates which are individually affected by several factors including race, gender, educational requirement, equal opportunity employer, and cities. Resumes with white sounding names have a 9.7 percent chance of receiving a callback, compared to resumes with African-American sounding names which have only a 6.4 percent chance of receiving a callback. Resumes with White names have a higher chance of getting a callback, which means that employers treat black people less favorably than white people. Candidates who send their resumes from Boston have a 9.7 percent chance of getting a callback. Equivalent resumes sent from Chicago receive a callback with just a 6.7 percent chance. There are differences in labor market conditions in Chicago and Boston. The gender differences at the workplace are quite significant. 8.2 percent of female names received a callback, compared to only 7.4 percent of male names. Educational requirements significantly influenced the callback rate. Positions requiring education are more competitive and selective than ones without education requirement.
Table 2: Mean callback rates by race on resume
Table 2 tabulates the average callback rates by race. This table shows that resumes with white sounding names always received a higher chance of getting a callback in terms of gender, education requirement, Equal opportunity employer, and city. Rows 2 and 3 display racial sounding resumes between male and female applicants. About 3 percent more callbacks were received by female and male applicants with white names than female and male applicants with African American names. There is a roughly 2 percent difference between male and female candidates. The numbers show that there is a significant racial gap for both males and females.
According to table 1 and 2, resumes with no education requirement are likely to receive more callbacks than resumes with education requirement, especially with white names. Resume with equal opportunity employer factor seems to have a slightly higher impact on callback rate than resumes without equal opportunity employer when applicants are differentiated by race. Like table 1, resumes sent to Boston have approximately 3 percent more callbacks than resumes sent to Chicago. Table 2 also measures the callbacks in these two cities but breaking down the resumes into White names and African American names. Unsurprisingly, callback rates are higher on resumes with White names than resumes with black names. The job market in Boston is less competitive than in Chicago. The results may suggest that there is a sexism issue in Chicago. Discrimination appears to vary a lot by city.
Table 3: Matrix of Coefficients for Models 1 - 3
Table 3 provides the coefficients and p-values of 3 different models analyzed in this paper. Model 1 includes explanatory variables of race and years of experience of the applicant, and a callback as a response variable. All three terms, including the intercept, were significant. However, this model received an AUC score of 0.5896, indicating that this model is not a very accurate predictor of the data within this dataset. Model 2 includes the race, years of experience of the applicant, and the city of the job posting as explanatory variables, with a callback as the response variable. All terms of this model were significant. This model received an AUC score of 0.604, which indicates this model is a better fit than model 1. Model 3 includes race, the city of the job posting, and educational requirement as explanatory variables, with a callback as the response variable. This model received an AUC score of 0.583, which indicates that it is not as good of a predictor of the data in this dataset compared to model 2.
Table 4: Models 4 and 5
Model 4 includes race, years of experience, city, educational requirement, and sex as explanatory variables, with a callback as the response variable. All coefficients in this model are significant. Model 4 received an AUC score of 0.608, which indicates that it is a better fit than models 1 through 3. Model 5 includes race, years of experience, city, educational requirement, sex, and number of jobs on each resume as explanatory variables, with a callback as the response variable. Model 5 received an AUC score of 0.6172, which indicates it is the best model out of the ones analyzed in this paper. This AUC score indicates that this is not extremely accurate for predicting the outcome of the information in this data set. Further study looking at a wider set of the variables could be done to better this model.
From model 5, the coefficients are exponentiated for interpretation. According to this model, a person with an African-American sounding name, with 0 years of experience, in the Boston area who is female and has 0 jobs on their resume has a 12.9% chance of getting a callback. These odds are increased by 1.56 if the applicant is white. Each year of experience increases the odds of a callback by 1.04. Applications in the city of Chicago have 0.64 times the odds of getting a callback compared to applications in Boston. If the job requires some level of education, the odds of a callback are 0.73 times lower. If the applicant is male, their odds of a callback are 0.77 times that of a female applicant. Each job included on the resume changes the odds by 0.886 times.
Discussion
From the result, we can conclude that African Americans have fewer opportunities in the labor market due to racial discrimination. There is a significant difference in terms of callback rates between job applicants with White names and job applicants with African American names. Callback rates are also affected by other factors, such as education, gender, years of experience, city (Boston and Chicago). An increase in callback rates happened when we combined these factors. This means that training programs may help to improve the callback rates for African Americans to some extent. However, we cannot deny the fact that racial discrimination is still a big issue at the workplace.
The exploratory data analysis gives answers for the questions above. Models show significant results about the average racial gap in callbacks. Our focus of these models is to support these two findings: African-Americans’ lower callback rates, and the effect of gender, city, years of experience, and level of experience on callback return rates. The models show that all of these factors are significant in explaining the changes of the callback rates.
The study of Ethnic Discrimination in the Labor Force offers an insight into African-Americans’ opportunities in the labor market. There is a significant difference in terms of callback rates between job applicants with White names and job applicants with African-American names. Callback rates are also affected by other factors, such as education, gender, years of experience, city [2]. An increase in callback rates happened when we combined these factors. This means that race is not the only factor that significantly influences the callback rate. The study helps us to understand the discrimination issue, and from that, we will know how and where to “fix” the problem. Thus, training programs are believed to be one of effective solutions which may help to improve the callback rates for African Americans to some extent [1].
The generalizability of our model includes applicants who apply to ads in the Boston and Chicago areas. We can only generalize to our sample population. Our model does not account for employers receiving so many resumes that when seeing an African-American name they disregard the entire resume, regardless of the skills listed. This weakness of our model suggests that perhaps other models may do a better job for explaining our findings. Another weakness of this study is that the names on resumes simply suggest race through personal names [1].
Some limitations of our study include the fact that some names may signal certain personality traits to employ. There is no way to guarantee that an employer will not infer social background from a name [1]. Some concerns for this limitation is the implication that employers discriminate not by the race implied from the name on the resume, but by the social background implied by the name [2]. This has the potential to impact the significance of race on callback rates. For future research, we could examine the relationship city with callback rates. Using two way tables and logistic regression we can find if city and sex or city and race have a relationship.Using employment fields we can determine if Boston or Chicago appears to have significant levels of racism and/or sexism.
Response to Peer Review
Big takeaways from the peer review for our group was to add pieces of analysis that we were missing. Some changes suggested from our peers that we made were adjusting our methods and materials to include data manipulation information and methods of statistical analysis. Along with that they suggested for results that we add summary statistics about the study population, for example percentage of white/black resumes sent out, so we added a plot comparing percentages and test statistics for each race of resumes. We initial had three tables, one for each variables, race, sex, and city. However our peers suggested we either join the tables together or get rid of them and since our research question is focused just on race we decided easy readability we would just use the tables significant in our analysis Lastly, for our discussion portion, they suggested more explanation and speculation on the interpretation of the data, as a result of this we were able to strengthen our report. Overall the peer review process was helpful for our final version of the report and without it we may have missed some important areas of conservation of our data.
Annotated Appendix
Summary Statistics
Table 1: Mean callback rates for all resumes
##
## No Callback Callback
## Black 0.9355 0.0645
## White 0.9035 0.0965##
## No Callback Callback
## Female 0.9175 0.0825
## Male 0.9262 0.0738##
## No Callback Callback
## Boston 0.9030 0.0970
## Chicago 0.9327 0.0673##
## No Callback Callback
## Not EOE 0.9200 0.0800
## EOE 0.9182 0.0818##
## No Callback Callback
## No Education Requirement 0.8977 0.1023
## Education Requirement 0.9254 0.0746Table 2: Mean callback rates by race on resume
##
## No Callback Callback
## White Female 0.9011 0.0989
## White Male 0.9113 0.0887##
## No Callback Callback
## Black Female 0.9337 0.0663
## Black Male 0.9417 0.0583##
## No Callback Callback
## Boston, White 0.8837 0.1163
## Chicago, White 0.9194 0.0806##
## No Callback Callback
## Boston, Black 0.9224 0.0776
## Chicago, Black 0.9460 0.0540##
## No Callback Callback
## Not EOE, White 0.9061 0.0939
## EOE, White 0.8970 0.1030##
## No Callback Callback
## Not EOE, Black 0.9340 0.0660
## EOE, Black 0.9394 0.0606##
## No Callback Callback
## No Education Requirement, White 0.8687 0.1313
## Education Requirement, White 0.9129 0.0871##
## No Callback Callback
## No Education Requirement, Black 0.9266 0.0734
## Education Requirement, Black 0.9379 0.0621Model 1: Callback by race and years experience
##
## Call:
## glm(formula = call ~ race2 + years_exp, family = binomial(link = "logit"),
## data = resume_tbl)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.7008 -0.4318 -0.4006 -0.3496 2.4537
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.998941 0.115612 -25.940 < 2e-16 ***
## race2 0.438409 0.107523 4.077 4.56e-05 ***
## years_exp 0.039089 0.009199 4.249 2.14e-05 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 2726.9 on 4869 degrees of freedom
## Residual deviance: 2693.2 on 4867 degrees of freedom
## AIC: 2699.2
##
## Number of Fisher Scoring iterations: 5## (Intercept) race2 years_exp
## 0.0498398 1.5502389 1.0398627Model 2: Callback by race, city, and years experience
##
## Call:
## glm(formula = call ~ race2 + city + years_exp, family = binomial(link = "logit"),
## data = resume_tbl)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.6409 -0.4394 -0.3900 -0.3307 2.4864
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.781565 0.133839 -20.783 < 2e-16 ***
## race2 0.439775 0.107627 4.086 4.39e-05 ***
## cityc -0.329331 0.108169 -3.045 0.002330 **
## years_exp 0.033201 0.009397 3.533 0.000411 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 2726.9 on 4869 degrees of freedom
## Residual deviance: 2684.0 on 4866 degrees of freedom
## AIC: 2692
##
## Number of Fisher Scoring iterations: 5## (Intercept) race2 cityc years_exp
## 0.06194148 1.55235864 0.71940476 1.03375788Model 3: Callback by race, city, and educational requirement
##
## Call:
## glm(formula = call ~ race2 + city + req, family = binomial(link = "logit"),
## data = resume_tbl)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.5391 -0.4381 -0.4031 -0.3258 2.4341
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.2954 0.1248 -18.395 < 2e-16 ***
## race2 0.4399 0.1075 4.091 4.3e-05 ***
## cityc -0.3522 0.1083 -3.251 0.00115 **
## req -0.2616 0.1227 -2.133 0.03296 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 2726.9 on 4869 degrees of freedom
## Residual deviance: 2691.3 on 4866 degrees of freedom
## AIC: 2699.3
##
## Number of Fisher Scoring iterations: 5## (Intercept) race2 cityc req
## 0.1007180 1.5525516 0.7031570 0.7698094Model 4: Callback by race, sex, educational requirement, city, and years experience
##
## Call:
## glm(formula = call ~ race2 + years_exp + city + req + sex, family = binomial(link = "logit"),
## data = resume_tbl)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.7409 -0.4371 -0.3884 -0.3275 2.5966
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.48550 0.16016 -15.519 < 2e-16 ***
## race2 0.44449 0.10777 4.124 3.72e-05 ***
## years_exp 0.03418 0.00950 3.598 0.000321 ***
## cityc -0.32538 0.11482 -2.834 0.004600 **
## req -0.32734 0.12449 -2.630 0.008551 **
## sexm -0.26643 0.13418 -1.986 0.047080 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 2726.9 on 4869 degrees of freedom
## Residual deviance: 2673.9 on 4864 degrees of freedom
## AIC: 2685.9
##
## Number of Fisher Scoring iterations: 5## (Intercept) race2 years_exp cityc req sexm
## 0.08328415 1.55968947 1.03476729 0.72225333 0.72083903 0.76611007Model 5: Callback by race, sex, educational requirement, city, years experience, and number of jobs on the resume
##
## Call:
## glm(formula = call ~ race2 + years_exp + city + req + sex + n_jobs,
## family = binomial(link = "logit"), data = resume_tbl)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.7430 -0.4381 -0.3844 -0.3239 2.6908
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.046619 0.236823 -8.642 < 2e-16 ***
## race2 0.445741 0.107836 4.133 3.57e-05 ***
## years_exp 0.040707 0.009811 4.149 3.34e-05 ***
## cityc -0.444481 0.123803 -3.590 0.00033 ***
## req -0.314923 0.124525 -2.529 0.01144 *
## sexm -0.266451 0.134168 -1.986 0.04704 *
## n_jobs -0.120688 0.049180 -2.454 0.01413 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 2726.9 on 4869 degrees of freedom
## Residual deviance: 2667.9 on 4863 degrees of freedom
## AIC: 2681.9
##
## Number of Fisher Scoring iterations: 5## (Intercept) race2 years_exp cityc req sexm
## 0.1291709 1.5616474 1.0415472 0.6411569 0.7298449 0.7660936
## n_jobs
## 0.8863106ROC Curve with AUC Measurement
## [1] 0.6172339Probability and Log Odds, Model G Setup
References
[1] Bertrand, Marianne, and Sendhil Mullainathan. 2004. “Are Emily and Greg More Employable Than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination.” American Economic Review, 94 (4): 991-1013
[2] Bursell, Moa. “What’s in a name?-A field experiment test for the existence of ethnic discrimination in the hiring process.” (2007).
[3] Cotton, J. L., O’Neill, B.,S., & Griffin, A. (2008). The “name game”: Affective and hiring reactions to first names. Journal of Managerial Psychology, 23(1), 18-39. doi:https://doi.org/10.1108/02683940810849648