Achieving Fairness in Data-Driven Decision Making

In insurance, fairness is an important concept since people are asked to pay different prices for coverage. But the principles are, of course, much broader.

For instance, actuaries are familiar with actuarial fairness—the principle that people with the same level of risk should pay the same price. Beyond pricing insurance risk transfers, actuaries are deeply involved in other aspects of a company’s decision-making processes.

This dual perspective—balancing company interests with societal implications—places actuaries at the heart of ensuring fairness in data-driven decision-making.

Fairness also includes community perspectives on customer fairness. The recent Actuaries Institute Dialogue Paper, Fairness in Insurance – A Challenge to Boards of Insurance Companies by Ian Laughlin calls on insurance boards to strengthen their oversight of customer fairness. It emphasises the need for regular evaluations of how a company’s products, pricing, claims processes and other practices align with community expectations.

In the realm of algorithmic bias and fairness, the focus often shifts to non-discrimination. An example of recent regulation is Colorado’s Senate Bill 21-169, which requires insurers to assess their algorithms and data to prevent unfair discrimination.

To strengthen this effort, the Colorado Division of Insurance proposed a quantitative testing framework in 2023 to identify unfairly discriminatory outcomes in algorithms and predictive models used for insurance underwriting. This highlights the growing emphasis on quantitative methods to achieve fairness in data-driven decision-making.

In this first article of my Responsible Data Science column, I delve into achieving fairness as a core principle of non-discriminatory decision-making in data-driven contexts. I look forward to exploring other aspects of responsible data-driven decision-making in future columns.

The myth of blindness (fairness through unawareness)

Imagine you’re an actuary pricing an insurance product or processing claims using data and algorithms. Your dataset includes both traditional and newer rating factors—perhaps from bank account data or telematics.

While there would be obvious privacy concerns to address too, let’s set them aside for now. Some anti-discrimination laws prohibit using protected attributes like gender, race or religion in pricing decisions.

But here’s a critical question: if you neither collect nor use these sensitive attributes (a concept known as “blindness or fairness through unawareness”), does that guarantee fairness?

The answer is no. Sensitive attributes are often associated with non-sensitive ones. For example, gender may correlate with vehicle engine size, and postcodes might reflect racial or ethnic distributions. These associations can result in unintentional discrimination.

In the era of AI and big data where vast datasets and complex algorithms are the norm, the risk of indirect discrimination increases. This is because as more information is added to the prediction, the more chance there is to have variables acting as proxies for sensitive ones. This becomes particularly concerning when there is no evidence of a causal link between these proxy variables and the actual risk or claims cost.

Without proper tools and awareness, companies and actuaries might miss these biases, perpetuating unfairness. Indeed, blindness has been the industry norm for ages; however, we need to be aware that discrimination may still exist even without using or collecting the protected attributes and we need to do something to mitigate it.

Start with a clear definition of fairness

The first and most challenging step in achieving fairness is defining it.

Fairness is context-dependent, varying by application area, line of businessand jurisdiction (Frees and Huang, 2021). Moreover, fairness definitions often conflict, making it impossible to satisfy multiple criteria simultaneously.

Below are some examples of fairness concepts that have been discussed in academia and industry, which has the potential to be applied to insurance contexts:

1. Fairness through awareness: Ensures similar policyholders are treated alike based on task-specific similarity metrics (Dwork et al. 2012).
2. Conditional demographic parity: Achieves statistical independence between decisions (predictions) and sensitive attributes after accounting for legitimate factors (Xin and Huang, 2023).
3. Equalised odds: Ensures that decisions (predictions) have equal true positive and false positive rates across different sensitive groups (Hardt et al. 2016).
4. Sufficiency (well-calibration): Ensures that decisions (predictions) are equally calibrated across sensitive groups (Baumann and Loi, 2023).
5. Counterfactual fairness: This criterion requires decisions (predictions) to remain unchanged in counterfactual scenarios where only the protected attribute is altered. It requires causal assumptions of rating factors (Kusner et al. 2017).
6. Controlling for sensitive attributes to avoid omitted variable bias: Recognises that excluding sensitive attributes can lead to biased outcomes due to unmeasured confounding and explicitly accounts for them by controlling for sensitive attributes in modelling (Pope and Sydnor 2011, Lindholm et al. 2022, Araiza Iturria et al. 2024).
7. Actuarial group fairness: The price charged on each sensitive group must be the same on expectation conditional on the estimated cost (Dolman and Semenovich, 2018).

Selecting the right fairness criterion is context-dependent and crucial, which requires stakeholder analysis and a thorough understanding of implications (Dolman and Semenovich, 2018, Shimao and Huang, 2022, Lindholm et al. 2024, Côté et al. 2024). It is not an easy task, but in my view, it is the crucial first step toward achieving fairness.

Once a notion of fairness is determined, we can then develop testing tools and mitigation strategies for assessment and implementation.

While regulations and laws specify protected attributes, there is a lack of clear guidance on addressing indirect discrimination or determining which fairness notions are most suitable in specific contexts.

Bridging this gap requires more research and collaboration between academics, policymakers and industry practitioners to develop actionable frameworks that can guide organisations in operationalising fairness in diverse contexts.

Focus on fairness in decision-making (beyond fairness in machine learning)

Data-driven decision-making processes are multi-step and complex. They involve algorithms (such as AI or machine learning) but also encompass other components.

In insurance pricing for example, fairness can be applied to various stages, including the pure premium, technical premium or market premium. While pure premium modelling (claims cost modelling) is typically considered a machine learning task, the broader pricing process involves additional considerations.

The above figure provides a general overview of the insurance pricing process commonly employed by practicing actuaries for certain lines of business.

Although cost modelling and demand modelling can be framed as statistical machine learning tasks, deriving the final pricing outputs involves additional steps, such as price optimisation^[1] and strategic pricing decisions (i.e., marketing and competition considerations). These aspects extend beyond the scope of a pure machine learning algorithm.

While most academic literature and recent industry discussions focus on algorithmic fairness or fair machine learning, real-world decision-making demands a broader perspective as achieving fairness in machine learning doesn’t guarantee fairness in outcomes.

For example, Shimao and Huang (2022) found that applying fairness-aware machine learning algorithms on cost modelling in an insurance context cannot achieve fairness in the market price or welfare alone. However, they can significantly harm the insurer’s profit and consumer welfare under certain market conditions, particularly of females.

Ultimately, fairness must extend beyond algorithms to encompass final decisions. Guidance like the Human Rights Commission and Actuaries Institute’s Guidance Resources: Artificial Intelligence and Discrimination in Insurance Pricing and Underwriting is invaluable, but more resources and discussions are needed to address fairness holistically.

To achieve fairness in decision-making, we need to consider the entire decision-making process, understand what the target of fairness is and what the final decision we care about entails.

The path to fairness

Fairness in data-driven decision-making is not just about algorithms; it’s about the broader impact of decisions on people and communities.

By understanding the limitations of blindness (or fairness through unawareness), clearly defining fairness notions, and considering the entire decision-making process, actuaries can lead the way in ensuring fairness in an increasingly complex, data-driven world.

There are many additional issues remain to be explored and resolved on the way to achieve fairness, which could serve as future topics for this column.

References

Araiza Iturria, C. A., Hardy, M., & Marriott, P. (2024). A Discrimination-Free Premium under a Causal Framework. North American Actuarial Journal, 28(4), 801–821. https://doi.org/10.1080/10920277.2023.2291524

Baumann, J., & Loi, M. (2023). Fairness and risk: an ethical argument for a group fairness definition insurers can use. Philosophy & Technology, 36(3), 45.

Charpentier, A. (2024). Insurance, biases, discrimination and fairness. Springer International Publishing AG.

Côté, O., Côté, P., & Charpentier, A. (2024) A fair price to pay: Exploiting causal graphs for fairness in insurance. Journal of Risk and Insurance. https://doi.org/10.1111/jori.12503

Dolman, C. and D. Semenovich (2018). Algorithmic fairness: Contemporary ideas in the insurance context. https://www.actuaries.org.uk/system/files/field/document/B9_Chris%20Dolman%20(paper).pdf.  

Dwork, C., Hardt, M., Pitassi, T., Reingold, O., & Zemel, R. (2012). Fairness through awareness. In Proceedings of the 3rd Innovations in theoretical computer science conference (p. 214–226). New York, NY, USA: Association for Computing Machinery. Retrieved from https://doi.org/10.1145/2090236.2090255

Frees, E. W. (Jed), & Huang, F. (2021). The Discriminating (Pricing) Actuary. North American Actuarial Journal, 27(1), 2–24. https://doi.org/10.1080/10920277.2021.1951296

Hardt, M., Price, E., & Srebro, N. (2016). Equality of opportunity in supervised learning. Advances in neural information processing systems, 29.

Kusner, M. J., Loftus, J., Russell, C., & Silva, R. (2017). Counterfactual fairness. Advances in neural information processing systems, 30.

Lindholm, M., Richman, R., Tsanakas, A., & Wuthrich, M. V. (2022). Discrimination-freeinsurance pricing. ASTIN Bulletin: The Journal of the IAA, 52(1), 55–89.

Lindholm, M., Richman, R., Tsanakas, A., & Wüthrich, M. V. (2024). What is fair? Proxy discrimination vs. demographic disparities in insurance pricing. Scandinavian Actuarial Journal, 2024(9), 935–970. https://doi.org/10.1080/03461238.2024.2364741

Pope, D. G., & Sydnor, J. R. (2011). Implementing anti-discrimination policies in statistical profiling models. American Economic Journal: Economic Policy, 3(3), 206-231.

Shimao, H., & Huang, F. (2022). Welfare cost of fair prediction and pricing in insurance market. UNSW Business School Research Paper Forthcoming.

Xin, X., & Huang, F. (2023). Antidiscrimination Insurance Pricing: Regulations, Fairness Criteria, and Models. North American Actuarial Journal, 28(2), 285–319. https://doi.org/10.1080/10920277.2023.2190528

Footnotes 

CPD: Actuaries Institute Members can claim two CPD points for every hour of reading articles on Actuaries Digital.