Image: Luca Bravo via Unsplash

Estimating and removing the gender pay gap: Analytic tools and their ethical dimensions

Swiss companies with more than 100 employees are obliged by law to perform a gender pay gap analysis. The interest in using analytic tools for salary data to identify, reduce and remove salary discrimination has increased. Important ethical questions arise when developing and implementing tools.

Not only since the Frauenstreik in 2019 has the gender pay gap become an object of public debate and a focus area in politics (in Switzerland, the gap amounted to an average of 18% in 2016, with half not explainable by factors such as age, rank, or education). Since July 2020, Swiss companies with more than 100 employees are obliged by law to perform a gender pay gap analysis. The interest in using analytic tools for salary data to identify, reduce and remove salary discrimination has increased. To support companies in this endeavour, the Swiss confederation developed Logib, an algorithmic analytic tool that is publicly available. Based on a ‘simple’ linear regression, Logib identifies the average gender pay gap within an organization.

The Challenge

While Logib gives broad insights on the company-wide size of the gender pay gap, many companies have encountered difficulties in reflecting their compensation structure in Logib. Additionally, Logib currently does not provide any actionable suggestions for HR or compensation specialists on how to remove the gender pay gap for individual employees. We at ethix - Lab for Innovation Ethics, together with our partners Swiss-SDI, were approached by a large Swiss company with over 15’000 employees to ethically verify an internally developed adaptation of Logib. Our clients decided for the external verification in order to identify potential conceptual or technical weaknesses. Ultimately, an independent review of the tool can help building trust and legitimacy among internal stakeholders and decision makers.

Adapting Logib to fit a specific firm

The tailored tool follows the same analytic structure as Logib, but includes two main adaptations: First, it adjusts the justified explanatory factors to the internal compensation structure and second, it extends the analysis of wage discrimination to other factors beyond gender. Besides gender, work-time percentage, nationality and region of employment were included as other potentially discriminatory factors. As a major addition to Logib’s company-wide identification and quantification of the gender pay gap, the adapted tool also aims to give specific indications for adjusting the salaries of particularly discriminated individuals. Specifically, it aims to identify which individual employees appear to be especially discriminated and provides suggestions of how much these individual salaries should be adjusted.

Our ethical verification approach

Our ethical verification methodology follows three steps: First, clarify the underlying ethical intent of the use case. Second, assess the development process and identify potential ethical risks associated with its use and third, verify that the technical translation (analytic structure, code) aligns with the ethical intent. As a result of our ethical verification process, we found that this additional dimension to the Logib analysis raises ethical and mathematical questions.

First, it is important to understand the underlying ethical logic of the Logib tool. Logib does not define what a “fair” salary might amount to or which criteria are acceptable or legitimate to determine salaries. Rather, Logib aims to prevent a specific type of discrimination, namely gender-based discrimination. It follows a negative definition of fairness, assuming that a salary is fair (other things being equal) if it is not gender-discriminatory. Second, from a statistical point of view, it is important to realize that discrimination detection and salary modelling are two very different goals for an algorithm. Whereas Logib was created to achieve the first goal, the adapted tool tries to achieve the second goal, while using an algorithm created to achieve something different. Just like Logib, the adapted algorithm only includes a selection of justified explanatory criteria. As a consequence, the tool risks identifying not the most discriminated individuals but those, whose salary is not well explained by the algorithm. Third, the Logib algorithm is designed to identify systematic discrimination within a group, not to measure individual discrimination. This finding also holds for the adapted tool. It may find that women earn 12% less on average across the entire company. This, however, does not mean that one particular female employee actually is discriminated against by 12% with regards to her direct male colleague.

Going Forward

We found that the tool developed by our clients cannot directly identify and quantify the most extreme cases of discrimination. Nonetheless, it is able to flag all employees, whose salary is inadequately explained by the model. The list of flagged individuals can serve as a preselection for an in-depth assessment that establishes if specific employees suffer from wage discrimination and if so, to estimate the necessary salary adjustment. Attributing power to human agents, such as managers or HR staff, for this case by case analysis is crucial. The combination of data analytics and human decision-making can prove to be an efficient strategy to eliminate certain cases of discrimination.

It is crucial that these human agents understand and trust the analysis performed by the adapted tool. Our independent and external verification is a powerful tool to build trust. With our joint expertise in data science and applied innovation ethics, ethix and Swiss-SDI are committed to work together to support Swiss companies to tackle the gender pay gap in their organizations by developing, checking and improving automated salary gap detection tools.