The actuaries were shown a map, old and arguably outdated: Omaha, Nebraska, divided into zones A, B, C and D.
But the story behind it was as fresh as the day’s news: racial discrimination.
Discrimination was present in the Omaha map as the C and D zones represented where Blacks lived primarily — and where the Federal Home Loan Corporation wouldn’t authorize lending — and where, decades ago, insurers would not write business. The map represents a practice now known as redlining.
“Areas that have been the site of racial discrimination are most vulnerable to economic and natural catastrophes,” said the map’s presenter, Birny Birnbaum, executive director of the Center for Economic Justice and a consumer liaison representative to the National Association of Insurance Commissioners.
Untangling and addressing that troubled past are challenges to today’s actuaries. They learned of several approaches to the problem in a session titled “Disparate Impact: The Impact of the Social Justice Movement on Insurance Rating” at the CAS 2021 Virtual Ratemaking, Product and Modeling Seminar.
The legacy of that map, Birnbaum said, lives on in today’s insurance rates, but not through intentional discrimination. “I believe there is very little of that in insurance today,” he said, but through proxy discrimination — the use of classification variables that serve as a proxy for race.
“This kind of unnecessary discrimination is amenable to your actuarial skill set,” Birnbaum said. The challenge is helping to develop rates that can address the issue while remaining neither inadequate nor excessive to all policyholders.
The CAS is actively addressing the issues of race and insurance pricing, noted moderator Mallika Bender, FCAS, co-chair of the CAS/SOA Joint Committee for Inclusion, Equity and Diversity. Last fall the CAS adopted an approach to race and insurance pricing with the goal of employing leadership, collaboration, research and education to let actuaries drive solutions in ways to benefit consumers and the industry.
Roosevelt Mosley, FCAS, CSPA, a principal and consulting actuary at Pinnacle Actuarial Resources, outlined several possible ways to examine price discrimination.
1. Do nothing.
This had been a standard industry response for decades but is not one anymore. As Birnbaum put it, “Following the murder of George Floyd, many insurance company CEOs made forceful statements … The place to start is an examination of their own company practices.”
2. Exclude certain risk classifications from rating plans.
Some states already do this with gender and insurance credit score, Mosley said. But it is hard to know if the strategy is effective. It might address the issue partially, but not completely, or not at all.
The tactic “is fairly direct,” he said. “It’s applicable across the industry … [but]there really isn’t a blanket solution that if you eliminated a variable or a group of variables, it would eliminate the problem.”
Do nothing … had been a standard industry response for decades but is not one anymore.
And there is an arbitrariness to the process of excluding a variable. “Does it really fix the problem?” Mosley said. “And where does the problem lie? Which particular elements should be excluded, and what are the criteria for an element making it onto the list or not?”
3. Introduce a variable that controls for a protected risk characteristic.
Birnbaum favors such a method and spelled out his vision of how it would work.
Picture an additive model with three risk characteristics:
Y = b0 + b1X1 + b2X2 + b3X3 + e.
To this you would add a control variable for race. Call it C1. The new model is:
Y = b0 + b1X1 + b2X2 + b3 X3 + b4C1 + e.
If any of the variables are correlated with race, the model will reveal it, Birnbaum said. Suppose X1 is perfectly correlated with race; adding the race variable will eliminate X1’s predictive power. If X1 has predictive power beyond race, that will make it a sharper variable.
Mosley noted such a control could systematically control for proxy discrimination. But applying the methodology to nonlinear models is more complicated.
4. Evaluate the final price’s impact on protected classes.
Here, prices would be examined to determine if the entire rating plan has disparate impact. If so, the discrimination would have to be addressed.
This would shift emphasis. Instead of evaluating rating variables as inputs, it would regulate the output of the entire rating plan, Mosley said.
If the plan is discriminatory, “there is a problem that needs to be addressed,” he said. It would also require collection of data from protected classes.
In general, Mosley said, bias can be addressed at three stages of model building:
- Removing data bias. “If the data itself is systematically biased, there are ways … to debias that data up front, so that you are introducing into the modeling process adjusted information that is not biased.”
- Modeling approaches. Some models can “adjust model parameters to eliminate bias by satisfying defined fairness criteria.”
- Adjusting for predictions. This is a process that explores “how you test outcomes for bias, and then correct for that bias.”
Birnbaum urged actuaries to confront the issue head-on.
“If it’s going to happen,” he said, “it’s going to happen through actuaries, leading” regulators and others.
The recording for the session, “Disparate Impact: The Impact of the Social Justice Movement on Insurance Rating,” is available for free here https://www.pathlms.com/cas/product_bundles/1938.
James P. Lynch, FCAS, is chief actuary and vice president of research and education for the Insurance Information Institute.