Remedying Contextual Biases in Recruitment: An Agent-Based Approach

How to cite

Ahmed, Tutan. 2025. "Remedying Contextual Biases in Recruitment: An Agent-Based Approach." AEA RCT Registry. AEARCTR-0015375. https://doi.org/10.1257/rct.15375-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Tutan Ahmed*

* Primary investigator

Registry ID
AEARCTR-0015375
DOI
https://doi.org/10.1257/rct.15375-1.0
First registered
Last updated
Status
In development
Trial dates
2025-03-03 – 2025-12-31
Countries
Private
JEL codes
J16, C93
Secondary IDs
IEC Approval: IIT/SRIC/DEAN/2024, Grant Sanction Order No: CRG/2023/006517
Registry record
http://www.socialscienceregistry.org/trials/15375

Abstract

It is well-known that individuals depend on heuristics and cues in decision-making since such shortcuts reduce the cognitive loads required in decision-making. These contextual factors can lead to biased and involuntary discriminatory behavior that goes beyond statistical or taste discrimination. We call this bias -"contextual bias". These contextual biases can take place due to the lack of sufficient in-group comparison among the different groups present in the recruitment. It can also happen due to the scale or the relative volume of the majority/minority candidates in the recruitment pool. Exploiting a multi-agent recruitment system, we examine whether such factors responsible for contextual bias formation could be addressed with a relatively simple intervention. We manipulate both “in-group comparison” and “scale” (or “relative volume”) of the candidates in the recruitment process. Both these factors are found to influence bias among recruiters. An important takeaway from the experiment is that contextual factors can lead to bias formation which is independent of the individual attributes of the agents or the job characteristics. More importantly, this contextual bias – could also be manipulated. There is a cumulative effect of such contextual bias when observed in a multi-agent recruitment system as the action of each agent provides the context for the next agents’ action. Due to the rooted nature of this bias to its context - this bias is without memory, i.e. there isn’t any cumulative effect when different rounds of hiring are considered.

Intervention

The main intervention involves creating an expanded pool of female candidates in the IT sector through a diversity initiative in collaboration with foundit. The additional pool of female candidate resumes will then be shared with the recruiters of our key partner intermediary hiring firm – Quess Corp IT vertical (QITS). The first node in the multi-agent recruitment process is the frontline recruiter of Quess Corp (henceforth D4) who retrieves resumes from multiple sources to line up the candidates.
D4/D3 team screens from the set of candidates procured from multiple sources interested in a particular assignment by looking at the match between the candidate profile and the job description. They observe the match and then call the candidate for a brief discussion. If the candidate agrees to proceed – the D4 will require the candidate to send relevant documents through email.
We use advanced proprietary software to obtain the match score of the candidates who are a part of the experiment to control for the match coefficient between the candidate profile and the job description. The D4 recruiter simply provides a list of candidates they screen from multiple sourcing channels. They don’t provide any score to any candidates. Our software – specifically built to rank candidates obtained from a variety of sources – then provides a rank to the candidate. This rank acts as a proxy for employers’ objective assessment during candidate callback.
The next node along the recruitment process is the screening of the resumes which are shared by the D4/D3 recruiters of Quess Corp by the internal recruiters of the client companies. Once this internal screening is done – the candidates are given a callback for interview. The interview is then conducted at the next node – which is essentially the internal recruiter of the client company with expertise in the domain.
We observe the decisions across subsequent agentic nodes – e.g. decisions to interview callback, decisions after the interview, whether offers are accepted, and even the extent of negotiations during the recruitment process. Essentially, we want to observe the role of a structural change – i.e. the change due to pool manipulation - on the cognitive level of the agents in the subsequent nodes and the consequent hiring outcomes.

Experimental design

The randomization is done at the job assignment level following a stratified randomization approach. Strata will be defined across verticals and cities. The D4 assigned with a particular treatment assignment will receive the additional pool of female resumes obtained from foundit. The D4 serving the control assignment will receive nothing. The D4s will keep on changing across assignments and the task allocation decision will be taken by the managers above D4. Therefore, the same D4 could be allocated to treatment or control randomly with assignments.
Job verticals which require specific female or male characteristics (e.g. males are required in Network Engineering jobs, and females are required in HR/Customer-centric jobs) are excluded from the experiment. Certain job assignments with explicit gender requests are excluded from the experiment. Again, the client companies with strong reputations for gender-based hiring (either proponents of diversity hiring or vice-versa) are excluded from the experiment. These steps are adopted to minimize the “explicit gender bias” and to minimize “statistical discrimination” in recruitment.
We have chosen a sector, i.e. medium and high-skill IT sector, where the skills are gender neutral and the hiring is based on ability. Moreover, the above-mentioned exclusion criteria are adopted to ensure that the skill requirements for the jobs are independent of gender characteristics. Therefore the male and female candidates applying for the jobs are identical in terms of their observed and unobserved skill set. We can argue that male and female candidates are picked from the identical qualification distribution – meaning the candidates acquire the skills, and expertise through similar formal education, certifications, and work experience and these are not gender-restricted. We claim that any differences in gender-based hiring must be due to contextual bias – as it is not due to statistical or taste discrimination which are taken into account by design.

Design details

Randomization method
Quess ITS operates a robust staffing system serving over 300 clients across 10 verticals, including Telecom, Auto, Healthcare, ITES, Digital, Retail etc. The recruitment process is managed through two units cutting across all the verticals under the Quess ITS: Accounts (A) and Delivery (D). The Accounts team liaises directly with clients, handling job requisitions and salary negotiations, while the Delivery team sources candidates efficiently through recruiters (D4) using the Quess ATS system. Key concerns in this system are minimizing Turnaround Time (TAT) and optimizing the candidate sourcing process. For the intervention, the experiment team is directly going to work with the delivery team. The intervention will be randomized at the job assignment level using a stratified randomization approach. Assignments and the corresponding D4 recruiters will be stratified by job verticals to ensure balance. All the current jobs posted every day are updated on the ATS by the backend team of QITS. The experiment team will be provided with “view” access and “assignment” access. A designated Project Associate (PA) from the team will allocate the assignments across the verticals into treatment and control groups respectively. Several Junior Project Assistants (JPA) from the experiment team will then share eligible female candidate CVs from the Triumph database with the treatment group D4s for the respective assignments. The control group will continue sourcing candidates through existing channels. The exclusive access of the Triumph database will be provided to the experiment team and not to any of the D4 recruiters. Stratification will be performed to evenly distribute assignments across different job verticals under the experiment. Therefore, for each vertical, it is expected that half of the assignments will be in the treatment group whereas the other half will be in the control group. It may be noted that we are not directly randomizing the recruiters. The randomization is applied at the job assignment level, meaning that each job assignment is independently assigned to either the treatment or control group. Because recruiters (D4s) are responsible for managing these job assignments, some recruiters may end up handling assignments in both the treatment and control groups. This means that recruiters themselves are not being randomized, but rather, the assignments they work on are randomly allocated to either receive the intervention (access to Triumph resumes) or not. Also, it may be noted that the client companies are not being randomized. Client companies will receive candidates based on whether their job assignments were randomly allocated to the treatment or control group. The client companies will make hiring decisions as usual, without any influence from the randomization process. They will have no information about the experiment during the period of the experiment.
Randomization unit
Randomization is done at the job assignment level. The number of vacancies varies across the job assignments.
Clusters
We are expecting to cater to 5000-5500 such job assignments during the period of the study. Monthly number of vacancies received is around 700-800. Assuming the intervention to continue for 7 months, we will have a total number of vacancies between 5000-5500.
Observations
Usually, 5-10 candidates are lined up for each vacancy. Considering each job assignment with 2 vacancies on average, and 5 candidates are lined up for each vacancy; we will have a total of 5000X2X5 = 50,000 candidates to be allocated into treatment/ control clusters during the experiment.
Arms
Treatment Group: 2,500 assignments (50% of the total) will be allocated to the treatment group. Recruiters in this group will receive access to the Triumph program resumes in addition to their usual sourcing channels. Control Group: 2,500 assignments (50% of the total) will remain in the control group, where recruiters will continue to source resumes using only the existing channels (Monster, Shine, Turnkey Suppliers, etc.). The treatment and control assignments will be evenly distributed across the verticals where gender sensitivity is not a major factor. These verticals include: Telecom, Auto, Healthcare, Hitech, Digital and Retail. Whereas ITES (is known for a female preference, primarily in call centre roles), QIMS (prefers male candidates for integrated managed services), and Desktop Support (prefers male candidates due to task-specific requirements).
Primary outcomes
Primary Outcomes: • Number of male/female resumes viewed and shortlisted • Male/Female interview attendance and success rates • Male/Female job acceptance • final salary offer to Male/Female candidates
Sponsors
Name: Science & Engineering Research Board (now ANRF), Department of Science & Technology, Government of India Location: New Delhi, India Url: https://serb.gov.in/

Cited by

No citation data available for this registration yet.

Keywords