Statistical Discrimination in Labor Markets: Microeconomics

Ban the Box policies, which remove criminal record questions from initial job applications, have not achieved their intended goals and may have worsened employment outcomes for disadvantaged groups. Research shows these policies increased racial disparities in employment for young Black men because employers cannot distinguish between applicants with and without records, leading them to statistically discriminate based on race. Additionally, employers still conduct background checks later, meaning people with records still face rejection. This illustrates how well-intentioned policies can produce unintended negative consequences.

Multiple studies provide evidence on ban the box policy effectiveness. Aegon and Starr's field experiment in New Jersey and NYC found that after ban the box implementation, racial disparities in employment callbacks widened six-fold, with white males benefiting most while black males saw decreased callbacks. Doleac and Hansen's quasi-experimental study using CPS data found ban the box adoption correlated with 3.4 percentage point employment decreases for black males and 2.3 percentage point decreases for Hispanic males, with no effect on white males. Additional evidence from drug testing and credit check bans supports the statistical discrimination hypothesis, showing that restricting information access can reduce employment for disadvantaged groups.

Professional credentialing systems can institutionalize discrimination when they require disclosure of demographic information. The Writers Guild of America requiring members to report sex, race, religion, gender identity, and disability status enables systemic discrimination against certain groups while favoring others based on identity rather than ability.

Fear drives discrimination against ex-offenders, with 50% of employers refusing to consider them and callbacks dropping 50% for those with records. Ban the Box reforms delay criminal record questions until after conditional offers, allowing informed assessments of context and rehabilitation. Durham, NC saw applicant numbers nearly triple after implementation, with 96.8% of qualified candidates receiving offers. Georgia adopted similar policies citing expanded talent pools and improved hiring outcomes. Jacksonville extended protections to city contractors, recognizing that meaningful reintegration requires systemic policy changes that address employer fears while expanding economic opportunity.

Ban the Box is a policy requiring employers to remove the criminal record checkbox from job applications. California and many cities have passed such laws. The policy aims to help African American males who have been disproportionately affected by high incarceration rates. The idea is to give applicants a fair chance before revealing criminal history.

Job application screening software often contains significant biases against protected groups. Research shows that women, people of color, and individuals with disabilities are disproportionately rejected during initial resume screening. Even when companies attempt to de-bias their software through machine learning, the algorithms can become too sophisticated and find alternative indicators of protected status—for example, identifying candidates who participated in women's sports teams and subsequently rejecting their applications.

Statistical bias in automated recruitment systems is insidious and accumulates gradually over time. When systems are used to screen large numbers of candidates, small biases in the algorithm can compound and create systematic discrimination. Organizations can detect these biases by periodically comparing AI screening results with human screening results, allowing them to identify and correct problematic patterns before they become entrenched. This ongoing monitoring is essential for maintaining fair hiring practices.

Automated resume-screening software can perpetuate and even amplify existing biases in hiring practices. Studies have demonstrated that identical resumes with different names receive dramatically different responses from automated systems, with names associated with certain demographics being filtered out before human review. This creates a self-reinforcing cycle where biased algorithms reduce the pool of qualified candidates from underrepresented groups, making it appear that fewer qualified individuals exist when in fact the filtering system itself is excluding them. Addressing this requires auditing and reprogramming recruitment algorithms to focus on relevant qualifications rather than demographic proxies.

In the United States and many Latin American countries, automated hiring systems analyze resumes. The ideal employee profile is created using historical hiring data that contains biases (gender ratios, class backgrounds, origins, ages). The algorithm learns these biases and amplifies them, creating systems that appear objective because machines make decisions but actually reproduce discrimination at an industrial scale. This represents new forms of dehumanization.

Algorithmic hiring tools promise efficiency and diversity but often fail. Research shows they reinforce historical discrimination: the 2015 AD Fisher study found high-paying job ads shown more to men than women despite laws against gender-based advertising. Amazon's AI recruiting tool, designed to increase diversity, learned to prefer male candidates and penalized resumes with 'women's' and graduates of women's colleges. Other examples include discrimination against people with mental illness and showing criminal record suggestions more frequently for African-American names. These tools often fail because they pick up signal like disability status or irrelevant attributes, and their effectiveness should be tested using the scientific method with falsifiable hypotheses.

Correspondence studies (also called audit studies) are a field experiment methodology used to detect discrimination in labor markets. In these studies, researchers create identical fake resumes and cover letters that differ only by one characteristic (such as race, gender, or age), then apply them to real job postings. The researchers track callback rates to determine whether employers respond differently to applicants based on the characteristic being studied. This method allows researchers to isolate the effect of discrimination while controlling for other variables that might influence hiring decisions.

Audit studies use experimental methods to detect discrimination by sending identical applications with only one variable changed (such as names associated with different racial groups). A landmark study created resumes with either traditionally African-American sounding names (like Jamal or Lakisha) or traditionally white-sounding names (like Emily or Greg), then applied for real job postings. Callback rates were significantly lower for resumes with African-American sounding names (6.45 callbacks per 100 applications versus 9.6 for white-sounding names). This method allows researchers to isolate the effect of race on hiring decisions while controlling for other resume characteristics.

The resume correspondence study design measures discrimination by sending fake resumes with distinctively black and white names to real job openings. The hierarchical structure includes: 108 Fortune 500 companies, each with up to 125 jobs across counties, with 8 applications per race type. A hierarchical potential outcomes model defines delta_jf as job-firm treatment effects, with firm-level delta_f as average across jobs. Random sampling and assignment make delta_f unbiased estimates of firm-level discrimination. Standard errors account for binary outcomes and sample sizes.

This comprehensive section covers three classic methods for measuring discrimination: surveys and interviews (suffering from social desirability bias and lack of variation), secondary data residual analysis (cannot isolate discrimination from other factors due to omitted variable bias), and correspondence audits (field experiments that randomize characteristics to observe actual behavior). The speaker defines discrimination as differential treatment on the basis of race that disadvantages a racial group. Race is typically signaled through names, but there is considerable variation—white names received response rates ranging from 6.2% to 16%, and if researchers had selected names that clustered together, they might have found statistically indistinguishable response rates between races. A meta-analysis of about 60 correspondence audits since 2004 found that employment discrimination against blacks is significantly higher than in housing, public services, or other contexts. About 15% of African Americans in New York State birth records have racially distinguishable names, meaning audits capture a specific version of discrimination that doesn't account for people who can pass as white.

Audit studies and correspondence studies are empirical frameworks used to identify discrimination. These methods originated in sociology but have been widely adopted in economics. They involve presenting identical materials (such as resumes) with only demographic characteristics changed to observe differential treatment. While these methods excel at identifying the existence of discrimination, they face significant challenges in identifying the underlying sources or mechanisms driving discriminatory behavior.