Avoiding biases and adverse impact in predictive hiring

Part of our job here in the workforce science team is to keep up to date with new research in Organisational Psychology. This might sound boring to some people – but we love it!

As massive nerds, we find nothing more exciting than seeing new progress in our field. This time, our knowledge-cravings took us all the way from Melbourne to Orlando, Florida, to this year’s SIOP conference.

The issue of Adverse Impact

An important issue within our field – and within the US in general – is adverse impact and hiring for diversity.

We are passionate about ensuring people are not discriminated against in selection methods, whether it is because of gender, age, ethnic background or sexual orientation.

(Actually, this is also one of the key values and driving forces behind why Paul, our CEO, founded Sapia.)

One key topic at this year’s conference was the combination of data science and behavioural science. Specifically, there were a lot of discussions around how these sciences can work together to minimise bias and discrimination in the hiring process.

The standard recruitment selection process

To give you some background as to why this is important, let’s explore what a standard selection process might look like today.

If you ever have applied for a job, it is likely you have gone through a process involving;

  • your resume
  • a cover letter
  • a psychometric test (personality/intelligence)
  • an interview
  • reference checks

As mentioned, pretty standard. This is typically the different pieces of information that recruiters would use to assess your suitability for a role.

However, from an adverse impact perspective, this isn’t good enough.

The reason is that humans are biased (there are a plethora of studies out there proving this). And even if our biases (in most cases) are unconscious, we still base discriminatory decisions on them.

Does your name predict future performance?

A research study by The Ladders found that recruiters only spend about 6 seconds looking at a resume. Using gaze-tracking technology they identified that recruiters spend almost 80% of this time on only a few items:

  • name
  • current title/company
  • previous title/company
  • previous position start and end dates
  • current position start and end dates
  • education

To most people that would seem reasonable. Our previous professional and educational experience should be predictive of future performance, right?

If you agree, it might surprise you that past job experience only has a 0.13 validity when used to predict performance (and your name certainly has nothing to do with how you would perform).

So not only is the information recruiters look at not actually predictive of performance, but it also has the potential to adversely impact minorities.

An eye-opening example

In the 1970s, the Toronto Symphony Orchestra was composed of almost all white males. A few years later, they caught on to their diversity issue and decided to do something about it.

One initiative was to introduce ‘blind auditions’. Individuals would perform from behind a screen, making the assessors ‘blind’ to who was performing. This meant that the performance was in the center of the assessment, not the individual.

The result?

The proportion of women within the orchestra increased from 5% to 35%.

Individuals within racial minority groups are also discriminated against based on resumes.

Research found that applicants with ‘traditional’ english names received an interview for every 1/10 resumes sent out. This is in contrast to applicants with African-American names, who only got an interview for every 1/15 resumes.

As the resume is one of the most common determinators of whether an applicant progresses to the next stage – it is alarming that this method can adversely impact minority groups.

Luckily, some progress is definitely being made to combat this.

Different techniques, for example blind recruitment, are increasing in popularity. Some progressive businesses have leap-frogged and started using artificial intelligence (AI) driven algorithms as a first step in their assessment process.

Important to keep in mind with AI and Adverse Impact

When using AI, it is very important to understand that the data put into the algorithm is of great importance. AI is often touted as the solution to the biases inherent in our thinking, but if not executed properly, AI can also become biased.

This is because an AI algorithm is only ever as bias-free as the data we used to build it.

It can be difficult to make sure AI is increasing diversity, and at the same time maintaining its predictive power. The predictive power is basically how good a model is at predicting good performance – and weeding out those who wouldn’t do so well.

To ensure best chance of success it is crucial that the data we put into AI recruitment tools is bias free.

One way is to control what you put into your AI models. Big Data can for example be dangerous, as it looks at adding all possible data sources of information to predict performance.

This could mean that the AI model learns that ethnic background is a predictor for success, which we clearly want to avoid.

To combat this issue at Sapia, we make the following decisions:

Targeted variables:

  • we only choose variables that minority groups do not answer differently to other groups

(if we did the model could learn to discriminate against these groups if the variable was considered predictive)

Test our predictors:

  • Are they in fact adversely impacting anyone?
  • Conduct adverse impact studies

When considering a new assessment tool, you should always ask your test provider the following;

How do you ensure the assessment isn’t biased against any gender, age or racial category, whilst remaining highly predictive of performance?

If they can’t give you a satisfying answer, it is definitely worthwhile considering another vendor.

Liked what you read? For further reading on how we minimise bias in our algorithms, head here.

Sign up to our newsletter