Skip to main content
SpringerLink
Log in

The effect of the 2016 United States presidential election on employment discrimination

  • Original Paper
  • Published:
Journal of Population Economics Aims and scope Submit manuscript

Abstract

We examine whether employment discrimination increased after the 2016 presidential election in the United States. We submitted fictitious applications to publicly advertised positions using resumes that are manipulated on perceived race and ethnicity (Somali American, African American, and white American). Prior to the 2016 election, employers contacted Somali American applicants slightly less than white applicants but more than African American applicants. After the election, the difference between white and Somali American applicants increased by 8 percentage points. The increased discrimination predominantly occurred in occupations involving interaction with customers. We continued data collection from July 2017 to March 2018 to test for seasonality in discrimination; there was no substantial increase in discrimination after the 2017 local election.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Log in via an institution

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7
Fig. 8
Fig. 9
Fig. 10
Fig. 11
Fig. 12
Fig. 13
Fig. 14
Fig. 15

Similar content being viewed by others

Data Availability

Data and other materials are available at Marina Gorzig’s website: https://sites.google.com/site/marinagorsuch/research?authuser=0

Notes

  1. In this context, we are using the traditional psychological definition of “salience”—the salience of a person, object, or characteristic is defined as how much it draws attention or stands out among its neighbors. Chetty et al. (2009) use a different definition of “salience,” where it refers to how visible a tax-inclusive price is. In that paper, a tax was more salient when the posted price included the tax, rather than it being added at the register. This is related to, but different from, the idea of salience as drawing attention to a particular person, object, or characteristic.

  2. In this context, “African American” is used to refer to black Americans whose families have been in the U.S. for multiple generations. While first- and second-generation immigrants from Africa may also identify or be identified as “African Americans,” we are using this term to refer more specifically to multi-generational black Americans.

  3. In this context, Somali is defined as having at least one of the following applies:

    1. 1.

      Answering “Somalian” as either the first or second answer to “What is this person’s ancestry or ethnic origin?”

    2. 2.

      Reporting being born in Somalia

    3. 3.

      Having a mother or father in the household who reports “Somalian” as their ancestry

    4. 4.

      Having a mother or father in the household who reports being born in Somalia

    This expansive definition is used for two reasons. First, some Somali Americans either do not report their ancestry or report it as African, East African, African American, or similar broad option. Using a more expansive definition will capture some of these people. Second, there is a persistent pattern where “Somalian” appears to be occasionally mistranscribed as “Samoan” in the ancestry question. The more expansive definition will count these people, if their birthplace, parents’ ancestry, or parents’ birthplace was reported correctly.

  4. Note that these first names are from the Koran and are not specific to Somali Americans. However, in the Minnesotan context, Somali Americans are the largest, most visible Muslim group.

  5. To the best of our knowledge, names from Minnesotan birth certificates are not public, so we are not able to use Minnesotan birth certificates to select names. We were unable to pretest Somali American names with a similar method because we would need to recruit a sample from only Minneapolis and St. Paul, which was cost-prohibitive.

  6. At the annual APPAM meeting, our discussant highlighted that most real job applicants do not have racially distinctive names. That is, most African American job applicants do not have names that identify them as African American. Similarly, the white names are Anglophonic—many real white applicants have names or surnames that are not from an Anglo-Saxon origin. This means that the results of resume correspondence studies do not generalize to the broader population of job applicants.

  7. We intended to have an equal number of resumes with generic, religious, and political activities. However, initially there was an error in generating the resumes such that 50% of resumes had a generic activity, 25% had religious, and 25% had political activities. This was corrected starting with resumes in August 2016.

  8. The Somali American refugee population is almost entirely Sunni Muslim (IIMN 2020). The impact of being Somali American will always include the impact of being Muslim—these two elements are not separable in reality or in the experiment. When a Somali American resume is randomly selected to include a religious activity, this can be interpreted as a signal of religiosity rather than religion. We do not include church activities on Somali American resumes because the effect would be difficult to interpret. Employers would likely view the applicant as a convert. We randomly select between church and mosque on African American resumes that are selected to include a religious activity.

  9. There are many jobs that our fictitious applicants are simply not qualified for, such as truck driving or healthcare positions that require a specific license. None of our resumes have these types of occupation-specific licenses, so we do not apply for these jobs.

  10. Probit model is presented in Gorzig and Rho (2020).

  11. It is important to note that we only know when the application was sent and when an employer contacted the fictitious applicant. We do not when the employer evaluated the application. Some applications that were sent just prior to the election may have been evaluated after the election. If this occurs, it will bias our findings toward zero. Those applications that were sent on election day itself are coded as “After the election” because most applications were sent in the evening and it is unlikely the employer read the applications immediately.

  12. This measure answers the question “How important is it to work with external customers or the public in this job?” and can be downloaded here: https://www.onetonline.org/find/descriptor/result/4.C.1.b.1.f?a=1

  13. The total number of applications that were sent by month is presented in Gorzig and Rho (2020).

  14. Regression results in which we include a time trend in the regressions with occupation fixed effects and no fixed effects can be found in Gorzig and Rho (2020).

  15. In fact, the sign of the correct marginal effect can be different for different observations. Because statistical software used to compute the marginal effects ignores this, traditional computations of the marginal effect of interaction terms in nonlinear models can result in incorrect estimates.

  16. We replicate our findings with a probit. These can be found in Gorzig and Rho (2020).

  17. The O*NET structure includes 964 detailed occupations. The RAs both coded each job in our sample with an occupation code. The two RAs agreed on the exact occupation for 74.5% of jobs. If the RAs coded the same job with two different occupations, we used the average of the “Deal with external customer” score from the two coded occupations. For example, one job was coded as “Cashier” by one RA and as “Retail salesperson” by the other. These occupations have scores of 91 and 97, respectively. This job was given the average of 94.

    We also coded occupations with AutoCoder, a machine learning algorithm developed for the Department of Labor that assigns O*NET occupational codes to job descriptions. Unfortunately, about 24% of the jobs in our sample had a match score below 70. Scores of 70 or above are generally considered to be a good fit. Therefore, we do not use AutoCoder for our analysis.

  18. Common jobs from the first tercile include dishwasher, carwash worker, or working in construction. The second tercile includes jobs like being an administrative assistant, cook, and data entry. The third tercile includes jobs like baristas, retail salespeople, customer service representatives, and being a server.

  19. Because we include job fixed effects, we cannot include month indicators by themselves.

  20. This is not due to changes in the number of job listings, but rather to changes in the number of hours the RA could work.

  21. Gorzig and Rho (2020) shows the monthly unemployment rate by race for 2013, 2014, 2015, and 2016. There is no increase in the black unemployment rate or the immigrant unemployment rate in November, refuting the idea that there is typically a seasonal increase in discrimination. The spike in discrimination we observe in the correspondence study is specifically targeted (Somali Americans) and does not reflect some broader seasonality of discrimination.

  22. In Gorzig and Rho (2020), we present the number of resumes that were sent by month for the July 2017 to March 2018 period.

  23. A job posting’s occupation category was chosen by the employer. We combined food/beverage/hospitality with customer service because of the small sample size of customer service jobs. The other categories are accounting/finance, business/management, healthcare, human resources, legal/paralegal, manufacturing, marketing/advertising/public relations, real estate, sales, salon/spa/fitness, science/biotech, security, skilled trade/artisan, technical support, and transportation.

  24. This correction is described in detail in Neumark (2012). David Neumark generously shared his code.

References

Download references

Acknowledgements

This project would not have been possible without our amazing team of research assistants. We are incredibly grateful for the hard work of Pachia Xiong, Joshua Edelstein, Anna Starks, Jenna Czarnecki, Carina Anderson, Hannah Lucey, Jordan Ewings, Zach LeVene, Michael Gleason, Giang Nguyen, and Mumtas Mohamed.

We thank the Russell Sage Foundation, Minnesota Population Center, and University of St. Thomas for generously providing funding for this project. This project has also benefited from insightful feedback from many people, including participants at the Inequality and Methods workshop organized by Jack DeWaard and Audrey Dorélien, David Eisnitz, Matthew Kim, Caroline Krafft, Samuel Myers, Jay Pearson, Evan Roberts, Seth Sanders, Andrés Villarreal, and Kristine West, and Patrick Button. We also thank the editor, Klaus F. Zimmermann, and three anonymous referees for their suggestions and insight.

Funding

This study was funded by the Russell Sage Foundation (#93-18-02), Minnesota Population Center Proposal Development Mentorship Grant (no number provided), and University of St. Thomas Partnership in Learning Grant (no number provided).

Author information

Authors and Affiliations

  1. Department of Economics & Political Science, Saint Catherine University, Saint Paul, USA

    Marina Mileo Gorzig

  2. Department of Economics, University of St. Thomas, Saint Paul, USA

    Deborah Rho

Authors
  1. Marina Mileo Gorzig
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Deborah Rho
    View author publications

    You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to Deborah Rho.

Ethics declarations

Conflict of interest

The authors declare no competing interests.

Additional information

Responsible Editor: Klaus F. Zimmermann

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendices

Appendix 1. Balance of resume elements with respect to key manipulations

The work experiences included on the resumes are selected randomly. To check this, we regress an indicator variable for key groups (white American, African American, Somali American) on the full list of work experience indicator variables. The following table shows the p-value of the F-statistic for jointly testing if any of the coefficients are significantly different from zero. None of the p-values are below 0.1.

Table 12 Work experience balance
Full size table
Table 13 Address balance
Full size table
Table 14 Order of resumes sent balance
Full size table

We also use a chi-squared test to examine if the address and order the resumes were sent are all balanced with respect to the key groups. In all cases we fail to reject the null hypothesis that the manipulations are balanced across these elements.

Appendix 2. Pre-election trends

The regressions we use in this paper are variants on a classic difference-in-difference. An important assumption for difference-in-difference analysis is that the two groups have the same trend prior to the intervention. As shown in Fig. A1, both white American and Somali American resumes were being called back more over time prior to the election. There is a slight difference in trend, with the proportion of white Americans contacted increasing at a slightly faster rate. African Americans do not appear to have this upward trend.

Figure 16:
figure 16

Trends in proportion contacted by employer before election. n= 324 (white American), 324 (African American), 648 (Somali American)

Full size image

Appendix 3. Main results by industry

In the main text, we showed that jobs in the two highest terciles of customer interaction experienced the largest increase in discrimination after the election. In Table 15, we display results from estimating Eq. 1 separately for the top three industry categories—administrative/office, food/beverage/hospitality/customer service, and general labor.Footnote 23

For Somali Americans and African Americans, we see a statistically significant increase in discrimination after the election in customer-oriented jobs, while there is no evidence of an increase in discrimination in the other job categories considered for this group.

Table 15 Results of a of linear probability model testing differences in discrimination
Full size table
Table 16 Results of a of linear probability model testing differences in discrimination during 2017 election
Full size table
Table 17 Results of a of linear probability model with triple interaction
Full size table

Table 16 shows the same analysis during the election of 2017. We find no increase in discrimination after the 2017 election in any industry.

We further examine the increase in discrimination by including a triple interaction between the customer service score, the ethnicity indicators, and the after election indicator. Prior to the election, the customer service score was not associated with being called back for white American, Somali American, or African American applicants. While not statistically significant, the relationship between customer service orientation and being called back became negative for Somali American applicants after the election. The results in the tercile analysis suggest that the relationship is nonlinear, which is likely why these results are weaker than those found in the tercile analysis.

Appendix 4. Neumark correction

One important consideration when interpreting the results of an audit study is that while all observed characteristics are carefully controlled on the resumes, there may be unobserved characteristics that have different variances between the different groups. Resume correspondence studies are typically attempting to detect discrimination—this can be statistical discrimination based on different averages in unobserved characteristics and discrimination based on taste/prejudice. However, if the groups have different variances in unobserved characteristics, the regression coefficients may not reflect the underlying discrimination because they “cannot separately identify the effect of race and a difference in the variance of the unobservables” (Neumark 2012, page 1136; also see Heckman and Siegelman 1993; Heckman 1998; Neumark and Rich 2019). The impact of the second moment of unobservables “is an artifact of how a correspondence study is done—in particular, the standardization of applicants to particular, and similar, values of the observables, relative to the actual distribution of observables among real applicants” (Neumark 2012, page 1148). That is, the sign and magnitude of the difference in the proportion of the applicants contacted can be the result of the experimental design and not reflect actual discrimination. The Neumark bias correction directly addresses this problem and separates out the effect of race and the impact of the second moment of unobservables.

Neumark (2012) developed a method to use observed characteristics to correct the bias caused by differences in variances from unobserved characteristics. The bias correction uses variation in observed characteristics on the resume to decompose the total difference into the portion due to the “marginal effect of race through level” (first moment statistical discrimination and prejudice/taste for discrimination) and the portion due to variance in unobserved characteristics (which is an artifact of the study—how the experimental resumes landed among real resumes). Because Somali Americans and white American applicants likely have different variances of unobserved characteristics, we implement the Neumark correctionFootnote 24 to estimate an unbiased overall discrimination effect between white American and Somali American resumes. To implement this correction, the experimental design must have variation in at least one observed characteristic in the study that influences perceived productivity and a testable identifying assumption that this characteristic affects callbacks homogenously across races. The correction uses a heteroskedastic probit model to test if the ratio of the variances in unobserved characteristics differs between white and Somali American applicants.

In our study, we have multiple observed variables including education, managerial work experience, customer service orientation of the job, sex, formatting of the resume, and order the resumes were sent. In column 1 of Table 18, we show the Wald test statistics for testing that these variables affect callbacks homogenously for white American resumes relative to Somali American resumes. As shown in the “Test statistics” section of Table 18, we fail to reject the null hypothesis that these characteristics affect callbacks homogenously across races, meaning we can use these variables to correct the bias from unobserved characteristics.

We then decompose the total difference in callback rates into the marginal effect of race through level—the discrimination coming from the employers’ taste for discrimination or from first moment statistical discrimination—and the difference due to variance. We find that the corrected estimates of discrimination remain negative and of similar magnitude to the naïve estimate. Likewise, the marginal effect of race through level is consistently large and negative. As shown in Table 18, the corrected estimate of discrimination against Somali Americans is −0.098, with the largest component coming from differences in the level (e.g., taste-based discrimination or first moment statistical discrimination) rather than the variance of the unobserved characteristics. The marginal effect of race through level is −0.073, while the effect of the variance is only−0.024. That is, the discrimination we find is not being driven by variance in the unobserved characteristics.

Column 2 shows the result of the correction when estimated using only the variables that were statistically significant in the probit estimate of predicting being called back by the employer. This subset of variables includes the formatting of the resume, the order in which the resume was sent, and including a college degree on the resume. The results of the correction are similar when just using this subset of variables: the largest component of discrimination comes from differences in the level rather than the variance of the unobserved characteristics.

Table 18 Neumark correction
Full size table

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Gorzig, M.M., Rho, D. The effect of the 2016 United States presidential election on employment discrimination. J Popul Econ 35, 45–88 (2022). https://doi.org/10.1007/s00148-021-00837-2

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s00148-021-00837-2

Keywords

JEL classification

Search

Navigation