When Testing Life Insurance Models, What’s Fair About Fairness Metrics and Methods?

How do you measure fairness in the context of life insurance? As actuaries, underwriters and data scientists increasingly face the question: is this model fair? In this Quick Read, we highlight key considerations for assessing fairness in a risk-based system, where different definitions of fairness can lead to very different conclusions, and metrics alone rarely tell the full story.

What is fair in a life insurance context?

There are various analytical approaches to determine fairness, but the underlying objective is consistent: to ensure that observed group differences reflect genuine differences in risk and are not discriminatory.  This distinction is especially important in life insurance, where risk differentiation is fundamental to how underwriting and pricing work.

Fairness in this context can be viewed through two lenses, “equality” and “equity”.  While both are fair, they can lead to very different results. An equality-based approach treats all individuals the same, meaning regardless of circumstances, premiums would be the same for everyone. An equity-based approach, by contrast, is grounded in proportionality: individuals who carry higher risk pay more, and those who carry lower risk, pay less. Life insurance is inherently built on this second principle, which makes the choice of fairness definition consequential.

What concepts matter when measuring fairness?

Discussions of fairness in modeling are global. Regardless of source or location, most legislation tends to focus on a common set of concerns:

• Direct discrimination – based on a protected class (e.g., race, gender identity) is prohibited.
• Proxy discrimination – where a variable closely associated with a protected class act as indirect substitutes.
• Technical bias – differences in baseline predictions, such as how factors are weighted, or accuracy rates across populations, can cause a model to perform differently for different groups.
• Disproportionate outcomes – focusing solely on adverse impacts to a group, regardless of the underlying circumstances, we risk ignoring legitimate risk-based differences.

What is a more meaningful approach to fairness?

A more effective approach begins with model validity that ensures inputs and outputs are logically and empirically tied to the risk being measured. Without validity, fairness analysis has little meaning.

From there, fairness assessment requires deeper analysis. Exploratory diagnostics, model-based testing, and structured comparisons can help determine whether observed differences are justified, or whether they signal potential bias. The objective is not to eliminate all differences across groups, but to ensure they are grounded in legitimate risk factors rather than bias.

Commonly used fairness metrics generally fall into two categories: threshold-based metrics, such as the 4/5ths rule, and continuous metrics that assess differences across the full distribution.  Interpreting these metrics is rarely straightforward:

• Results can be sensitive to directionality, which could be better for declines than acceptances.
• They are difficult to use at a group level to tell if a disparity is a real risk difference or results from a bias.
• Binary cut-offs can materially affect conclusions, making results sensitive to where thresholds are set.

Further approaches, such as exploratory comparisons, model-based testing, and causal analysis are often needed to explain differences across groups and assess whether they are justified.

What does this mean for insurers?

Fairness cannot be reduced to a single metric. It requires a structured, context-aware approach that balances responsible data use with the principles of risk-based differentiation. For life insurers, this means preserving the integrity of underwriting while ensuring that model outcomes remain fair, transparent, and defensible.

If you are interested in learning more about fairness metrics and methods and their role in life insurance, please access our paper.

Contributors

Jody Daniel, Senior Data Scientist, Life & Health

Tom Fletcher, SVP, Global Head of Data Science Consulting, Life & Health

Opinions expressed herein are solely those of the authors. This paper is for general information, education, and discussion purposes only. It does not in any way constitute legal or professional advice and does not necessarily reflect, in whole or in part, any corporate position, opinion or view of PartnerRe or its affiliates.