All Articles PartnerRe Perspectives Previous Article

Ethical standards are an integral and critical consideration for predictive model developers and users

Dr. Tom Fletcher, VP Data Analytics, North America Life, shares his perspective on the component strands necessary to build and maintain ethical standards in predictive modeling.

Customer experience in the modern era is an imperative and is guided by 24/7 news coverage, social media and new expectations of shopping. Companies can live and die with tight margins based on brand perceptions and personal experiences. Expectations for speed of delivery, consistency and a near arms race for efficiencies have led to a myriad of opportunities for predictive modeling.

With predictive models, and key to customer experience, comes the responsibility of meeting ethical standards and minimizing adverse impact. The relevant standards are those found in many professional codes relating to common principles of objectivity, integrity and fairness. These principles are strongly associated with three basic concepts that developers and users of predictive models must consider: reliability, validity, and fairness.

With predictive models, and key to customer experience, comes the responsibility of meeting ethical standards and minimizing adverse impact.”

Reliability: Models should be developed on data that will be reliable and consistent over the lifespan of the model. A house should not be built on shifting sands. Having a consistent model is another avenue to enhancing the customer experience via brand as well as personal experience. Reliability is the precursor to validity.

Validity: Standards of excellence should apply when evaluating models. When and under what circumstances does a model predict well? It’s often stated that all models are wrong, but some are useful. While this is true, the degree of error should be known and quantified. One would not allow a medical test into practice without knowing the sensitivity and specificity of the test and how it adds value in relation to other known tests (valid and creates a new cost-effective measure). While models should be valid, they should also be deployed to ‘do no harm’. This principle of beneficence is directly related to the fairness of the model usage.

Fairness: To discuss fairness, one must disentangle technical bias from sociopolitical views of right and wrong. Models are meant to discriminate risk, opportunity, or other target outcomes, but should be relatively free from technical bias, a concept whereby a model does not operate (predict) similarly across known groups. Systematic over/underprediction for certain groups will impact the customer experience of those groups and render reputational harm unless such bias is fairly and justly addressed. Unfair discrimination via predictive models exists when persons of equal risk (opportunity) are disproportionately selected into varying risk (opportunity) categories. Put differently, if two individuals from two groups scored the same, had the same level of risk (or potential for success) yet were disproportionately selected, then an unfair situation exists and customer experience will suffer.

Adverse impact is the outcome of model usage, not the constituent parts of the model itself. In some disciplines, adverse impact analyses are mandated (e.g. in personnel selection via the Uniform Guidelines on Employee Selection Procedures by the federal government). In other areas, more proactive stances to avoid adverse impact must be taken to ensure and indeed enhance the customer experience.

Techniques likewise exist to reduce such adverse impacts even when justified. Rather than focus on the constituent elements of a predictive model, consider adding components that mitigate the potential impact. A battery of assessments is generally more valuable (if administered innocuously) than a single data point. Ensuring the use of models meets these essential criteria will lead to trust, brand enhancement and overall greater customer experience.

PartnerRe’s Analytics team of behavioral and data scientists employs these concepts in model development and provides clients with advice and advocacy regarding ethical model development, evaluation and usage.


Tom Fletcher, VP, Data Analytics, North America Life

This article was first published in LOMA’s RESOURCE magazine, issue September/October 2020.
Opinions expressed are those of the author. This article is for general information, education and discussion purposes only. It does not constitute legal or professional advice of PartnerRe or its affiliates.
Get in touch