Written by Nathan Hewitt
Identifying and mitigating ethnicity bias in structured interview responses
This research paper is part of our accepted submission to SIOP, and will be presented at the 2023 SIOP Conference in Boston.
Discrimination based on race and ethnicity in personnel selection is a well known and pervasive issue highlighted in numerous studies (Bertrand & Mullainathan, 2004; Kline et al., 2021; Pager et al., 2009). Most of these studies report name-based inference of race and ethnicity by human reviewers leading to differential outcomes in the recruitment process. Linguistic racism is a form of discrimination that occurs based on one’s use of language, especially English (De Costa, 2020), and is highly associated with race and ethnicity.
As machine learning models are adopted to automate tasks like interview scoring, race or ethnicity encoded signals in language can lead to biased outcomes, if not mitigated. Hence understanding the level of ethnicity encoded signals in language is important when building natural language-based machine learning models in order to avoid biased outcomes, for example by using feature scores rather than raw text to score responses (Jayaratne, Jayatilleke, Dai, 2022).
In this work, we sought to quantify and compare the amount of ethnicity encoded information in over 300,000 candidates’ raw text interview responses to language-derived feature scores, including personality, behavioral competencies, and communication skills.
First, we trained machine learning models to predict candidate ethnicity from raw-text chat interview responses. Specifically, we trained an Attention-Based Bidirectional Long Short-Term Memory (Attn-BiLSTM) (Zhou et al., 2016) model for predicting ethnicity from textual responses.
Secondly, we tested the same for the language-derived features used in the automated scoring of the interview responses. We trained multiple models using a variety of machine learning algorithms (a linear model, tree models with bagging and boosting, and a neural network model with a single hidden layer) suitable for tabular data for predicting ethnicity from the 21 derived features.
Each model was then used to predict ethnicity for the 10% of the sample left out of the training dataset. The results from the classification tasks show a clear distinction between the ability to infer ethnicity based on natural language and inferred features. As hypothesized, we found that features derived according to a clearly defined rubric contain significantly less ethnicity information compared to raw candidate responses. That is, the models based on derived features recorded consistently weaker accuracy, precision, recall, and F1 values across all models compared to the model for the raw text candidate responses.
This research demonstrates the benefit of using algorithmically derived feature values in mitigating ethnicity related biases when scoring structured interview responses. Specifically, our results show that natural language responses to interview questions carry higher amounts of ethnicity information compared to features derived according to a clearly defined rubric for assessing interview responses. This further strengthens the case for using structured interviews that have been shown to reduce bias over unstructured interviews (Levashina et al., 2014) with much stronger criterion validity (Sackett et al., 2021).
References:
Bertrand, M., & Mullainathan, S. (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.
De Costa, P. I. (2020). Linguistic racism: Its negative effects and why we need to contest it. International Journal of Bilingual Education and Bilingualism, 23(7), 833–837.
Jayaratne, M., Jayatilleke, B., & Yimeng Dai (2022). Identifying and mitigating gender bias in structured interview responses [Paper presentation]. 2022 Society for Industrial Organizational Psychology Conference. Seattle, Washington, United States.
Kline, P. M., Rose, E. K., & Walters, C. R. (2021). Systemic Discrimination Among Large U.S. Employers (NBER Working Papers No. 29053). National Bureau of Economic Research, Inc.
Levashina, J., Hartwell, C. J., Morgeson, F. P., & Campion, M. A. (2014). The Structured Employment Interview: Narrative and Quantitative Review of the Research Literature. Personnel Psychology, 67(1), 241–293.
Pager, D., Western, B., & Bonikowski, B. (2009). Discrimination in a Low-Wage Labor Market: A Field Experiment. American Sociological Review, 74(5), 777–799.
Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2021). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology.
Zhou, P., Shi, W., Tian, J., Qi, Z., Li, B., Hao, H., & Xu, B. (2016). Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), 207–212.
Related
Blog
Why video screening will kill your D&I star
Right now video screening is the solution of choice for many, given the challenges of recruiting during the pandemic. Every day I’m asked about video solutions, and every week there seems to be a new video solution for hiring.
This isn’t people simply switching to Zoom, but rather embracing AI video platforms where you are judged by algorithms. Often algorithms crawl these videos to identify top candidates. This is not great. In fact, it’s horrifying. Not all video interviews are bad, given the pandemic it’s often become a necessity as a default for face-to-face interviews in the final stages of a recruitment process. But when it comes to top-of-the-funnel screening with first interviews, video interviews lead to biased outcomes.
Put simply, image and video recognition is built to favour white faces. In the documentary Coded Bias an M.I.T. Media Lab researcher Joy Buolamwini found that the algorithm couldn’t detect her face–until she put on a white mask. There are hundreds of validated research findings which confirm this.
Video Screening
Video invites judgement. It adds stress to the candidate with added pressure around hair and makeup, picking the right fake backdrop (yes, there are hundreds of advice columns on this), and practising and rehearsing your answers until you nail the recording. It turns a simple interview into a small theatre production.
Not everyone is comfortable on video, most especially introverts, people with autism, and people who feel marginalised. These factors do not influence or speak to a person’s ability to do a job, but by using video as part of the interview process they are put at a deep disadvantage. What percentage of people are you excluding just by using video?
Chat is a better option. It solves the challenges of remote interviews while being inclusive.
Try it for yourself, we’ll send you real results.
Blog
The ROI of Hope | Candidate Experience
Getting your organisation’s candidate experience right is proving to be something we’ll hear about increasingly. This is as job applicants demand more from companies they interact with. A recent poll on LinkedIn by a recruitment specialist attracted almost 80,000 views and 1000’s of reactions. It revealed that 52% of the people polled (one assumes mostly recruiters) believe that a template email is good enough as a response to an application for a job.
When it comes to recruitment, responding to candidates has always been an area we know has been ripe for improvement. And it costs companies too, with a bad candidate experience said to have cost Virgin $5 million.
That’s not to say there aren’t historical reasons why recruiters have not been able to respond to the hundreds of applicants received for a job. Until now it’s not been something we can practically do with limited time and resources. This is where AI plays a fundamental role in moving our industry forward as it allows mass personalisation at scale.
Hiring With Heart
This has never been more important than right now as we have had mass job losses across industries due to the impact of COVID-19. If you look at the Hospitality and Tourism industry across the globe, it’s hard to wrap your head around the sheer number of job losses with very little hope of returning to normal soon. If with every job application we were able to give each unsuccessful candidate feedback on where they could improve, imagine the impact we could have in activating the world economy.
This is entirely possible and every day we hear about the impact our technology is having on people’s lives when they get personalised feedback designed to steer them in the right direction.
81% of people who get personalised feedback from our platform said it was useful in identifying their strengths, 71% said it would help them better prepare for interviews and 59% said it would help them find a job that suited them.
And lastly, it’s not something we can quantify, but we do believe it’s important. What we’re giving so many people right now is hope. We think that’s something worth companies cultivating alongside us too.
To keep up to date on all things “Hiring with Ai” subscribe to our blog!
Finally, you can try out Sapia’s Chat Interview right now, or leave us your details to get a personalised demo
Have you seen the 2020 Candidate Experience Playbook? Download it here.
Blog
Recruitment metrics: How and why to track your candidate abandonment rate
In the candidate short market we’re in, it’s absolutely critical to keep talent engaged throughout the entire application process. You simply cannot afford to lose the talent that you’ve spent time and money attracting.
This sounds obvious, of course, but abandonment is a key problem – and few companies know where, when, and why it is happening.
Let’s start with the metric, and then talk about how we apply it to your wider talent acquisition journey.
Overall candidate abandonment rate = number of candidates still in the process at shortlist stage, minus the total number of candidates who landed on your careers page, divided by that total number again. Or:
At the very minimum, this is the metric you need to start tracking, because it is a generalized diagnostic for the health of your recruitment process.
If you know that you had 100 visitors to your careers (or job ad) page, but your shortlist has only 10 candidates in it, you’ve lost 90% of your possible talent pool at one stage or another.
Simple math, yes, but in our experience, many recruiters and talent acquisition managers don’t look at what their starting pool of candidate interest was – and therefore, what their theoretical talent pool might have been – and look only at actual applicants.
This poses another, related question: How do I know what my abandonment rate is at each stage of the application process?
Let’s say, like the example above, that you had 100 visitors to your careers (or job ad) page, and 20 of them completed the first-step application form on that page. You’ve lost 80% of your possible pool right there.
Not great, but at least you know – now you can examine that page to uncover possible issues preventing conversion.
- Is the page too long?
- Does it have too much text?
- Is the ‘apply’ button clearly shown?
- Is the form too long, requiring too much information to fill out?
- Are your perks/EVP attributes clearly displayed?
Without examining stage progression in isolation, you might never know why people aren’t sticking around.
To reiterate: As well as an overall abandonment rate, you need to measure the drop out rates at each of the stages of your talent acquisition journey. The next section can help show you what to focus on.
Where, when and why do candidates drop out of the application process?
Conventional wisdom tells us that the longer your application and interview process goes on, the higher your dropout rate will be.
But that’s a generalized issue – it tells you nothing about how to fix the problem, beyond simply making it shorter. You need specific, localized data to diagnose and fix your leakage spots.
Data from a 2022 Aptitude Research report on key interviewing trends found that candidates tend to drop out at the following stages, in the following proportions:
- 22% of candidates drop out at the application stage
- 24% at the screening stage
- 25% at the interview stage
- 18% at the assessment stage
- 9% at the offer stage
Good to know, right? If you audit your own journey, looking at these stages and using these numbers as benchmarks, you can quickly identify your weak areas.
For example: You might be proud of your four-step culture-building interview process, in which candidates have a coffee meet-and-greet with the team they’re hoping to join.
But if it’s cumbersome for the applicant and relies on several stakeholders to orchestrate, it may be dragging your process out unnecessarily, and doing more harm than good.
The most common job application abandonment stage: The interview
25% of candidates drop out here. Shouldn’t really be a surprise, should it? Job interviews are long, numerous, and in many cases, ineffective. According to Aptitude Research, 33% of companies aren’t confident in how they interview; 50% believe they’ve lost talent due to poor interviewing.
When asked about their top interviewing challenges, surveyed HR and TA leaders responded:
- Our interview process is too long (52%)
- We make candidates go through too many interviews (39%)
- Out interview process is inconsistent (38%)
- We don’t use objective data to drive decisions (32%)
- We have bias in the interview process (21%)
Let’s focus on that second-last challenge: lack of objective data. Almost a third of companies are approaching their interview and application process with assumptions and gut feelings; and half of them believe their interview process is too long.
Despite this, 68% of companies say they have not made any improvements surrounding candidate experience this year. How many, then, are looking seriously at their entire talent acquisition journey to see where it’s failing?
This is why we’re focusing on candidate abandonment rate in this post: It is a simple metric to show the health of your application process, easier to measure than many of the other recruitment metrics for which you’re responsible (the ever-nebulous quality-of-hire being a prime example). As the saying goes, what gets measured, gets managed.
Start here today, and see what you learn.
(P.S. Sapia’s Ai Smart Chat Interviewer combines the first four stages of your process – application, screening, interviewing, and assessment – together, resulting in an application process that can secure top talent in as little as 24 hours.
Because it’s a chat-based interview with a smart little AI, your team doesn’t need to do anything – everyone who applies gets an interview, immediately. That maximizes your talent pool right from the get-go.
What’s more, our candidate dropout rate is just 15%, on average. That means that 85% of your starting talent pool will stick around.
Why do our candidates stick around? More than 90% of them love the experience. See how we can help you here, today.)