Here’s a data science framework to assess gender inequity in hiring

Gender inequity is an important topic in the tech industry, but well-designed studies based on rich data are scarce. Firms have abundant data but lack a rigorous scientific framework, thus preventing HR departments from understanding the root cause of such inequity.

The primary goal of this post is to propose a replicable framework and methodology to assess inequity in recruiting, with a case study to illustrate.

Methodology

Data

Our dataset contains 1,382 applicants—1,029 (74.4 percent) males, 353 (25.5 percent) females—that applied to a data engineering position.

The gender of each applicant is predicted using our gender prediction model, which uses information from the candidates’ resumes. The error rate of the model is 4 percent, which is taken into account for all analyses.

The applicant’s skills, on the other hand, are extracted using our skill mapper model.

Measure of inequity

A good way to identify the potential inequity between genders is by comparing their rejection rates for a specific position. If all things were equal between applicants from each group, so should the rejection rates.

In this study, we measure inequity by comparing rejection rates in the application review stage. The reasons we limited rejection rates to just this stage are two-fold:

1. There are many factors that go into assessing an applicant (e.g. communication skills during a phone screen) that can’t be assessed with just a resume. So, we focused on this stage to limit the number of variables.
2. The application review stage often has the largest impact in the hiring funnel. We find that approximately 90 percent of all applicants get rejected at this stage.

Any difference in rejection rates in this stage can then generally be attributed to either objective attributes, like years of experience, education, and skill sets, and subjective attributes like the “quality” of education and experience and unconscious biases.

Since subjective attributes are subjective in nature, we limited the controls to objective ones.

Testing for significance

Analysis

Discussion

Limitations