Sources of Discrimination, Inaccurate Beliefs, and the Role of Blind Hiring

Employers assigned to the first treatment group (i.e., Treatment 1) will receive a
set of 20 hypothetical resumes with the name of a job applicant revealed. However,
before they are told to rate each hypothetical resume, they will be informed about
the coding tasks that UM students took as well as about the distribution of students’
actual coding scores by gender and race.3 We hypothesize that such information
provision of the distribution will correct inaccurate beliefs that employers may hold
about the distribution of the coding ability by gender and race of the job applicants,
by providing accurate information on coding ability by each group. The estimates
obtained using this treatment group will be used to estimate the discrimination caused
by inaccurate beliefs as well as the discrimination caused by animus (see Section 3.2
for details about the decomposition).

Employers assigned to the second treatment group (i.e., Treatment 2) will receive the
same set of 20 hypothetical resumes and will initially rate the hypothetical resumes
with the name of a job applicant blinded. However, the names will be revealed when
employers reach to the “mock interview” stage of the experiment. This treatment
arm will be used to study how blind hiring affects employers’ evaluation of resumes
and to examine if the effect of blind hiring persists even after the name and gender
are revealed during the “mock interview’ stage. This is designed to mimic an actual
job application process where employers eventually observe the gender and race of
job applicants during the interview stage even if the name is initially blinded during
the resume-screening stage. More details on estimating the effect of Treatment 2 are
provided in Section 3.3.

Finally, employers assigned to the control group will receive the same 20 hypothetical
resumes but with the name of a job applicant revealed and without the information
on the distribution of students’ actual coding scores.

2024-03-21

2025-05-30

We consider the following three primary outcomes:
** Employers' subjective evaluation of the hypothetical resumes in terms of coding skills.
-This outcome pertains to the question that reads: ``How good do you think this person will be in coding? [scale from 1-10]". Employers will be asked to answer this question for each of 20 hypothetical resumes during the ``resume-screening" phase.
- This outcome will be used to estimate and decompose into discrimination caused by animus versus discrimination caused by inaccurate beliefs as well as to estimate the effects of Treatment 2 (see Sections \ref{Estimating Discrimination} - \ref{Treatment Effect of Blind Hiring}).
**Coding scores estimated based on actual student resumes collated via the UM Student survey.
- This outcome will be used to decompose discrimination into discrimination caused by animus versus discrimination caused by inaccurate beliefs.
**Predicted coding scores associated with each hypothetical resume
- This outcome will be used to estimate the treatment effect of blinding names on the productivity of resumes that proceed to the ``mock interview" stage as well as on whether the resume is chosen for the final hiring (see Section \ref{Treatment Effect of Blind Hiring}).

To test the predictions of our theoretical model, we designed a novel two-step experiment that allows us to: i) estimate the productivity of students via coding test; ii) estimate the preference of employers via Incentivized Resume Rating (IRR); iii) estimate the treatment effect of blind hiring on the average productivity of workers hired; and iv) decompose the discrimination into taste-based discrimination and statistical discrimination caused by inaccurate beliefs.

We run our experiment in the context of US tech industry, with the University of Michigan computer science junior and senior students as the job applicants. US tech industry is a good context to study discrimination because there is well-documented discrepancies in the representation of gender or race in the US tech industry. Census Bureau reports that women only comprises 21\% of the US tech industry, and White comprises 64\% of the US tech industry whereas Black only comprises of 4\% in 2021. In addition, employers in the tech industry typically give out `technical interview' that asks job applicants to perform some coding tasks to observe the applicants' productivity. This setting maps well to our model, with `technical interview' representing $\tilde{s}$ in our model.

Not available

Randomizaton done in a survey platform Qualtrics.

Individual

No

N/A

120 employers

33 employers for treatment 1, 33 employers for treatment 2, 34 employers for control