The Markup is all about holding accountable the technology that powers our lives. So, for our first story we investigated an algorithm that sets prices for a product you need for everyday life and you MUST purchase or you can land in jail – car insurance.
When most people think big tech, they don’t think about car insurance. But the insurance industry pioneered big data and algorithmic predictions long before anyone in Silicon Valley. And when it comes to algorithms that matter, car insurance is a big one. Drivers are legally required to have car insurance in virtually every state. For many people, not being able to drive means you can’t work, take your kids to school or go to the doctor.
Car insurance algorithms are supposed to set prices based on how risky a driver you are. But we found that Allstate has been calculating prices with a secret algorithm
that appears to let prices float higher if you’re unlikely to shop around for a better deal. Allstate calls it a customer “retention model.”
When we dug into the algorithm we found that it was essentially a suckers list that squeezed customers who were already paying the most.
And surprise, surprise, some groups were more likely to be deemed suckers than others. Those who would have received massive rate hikes under Allstate’s plan were disproportionately middle-aged (41-62), disproportionately male, and disproportionately living in communities that were more than 75 percent “nonwhite,” according to census data.
Allstate has countered that our work is “inaccurate and misleading” because the ratings plan that we analyzed never went into effect. It was ultimately rejected by Maryland as “unfairly discriminatory.” But Allstate declined to tell us why our findings are not relevant for the ten states that have approved its ratings plans that mention using a retention model: Arizona, Arkansas, Illinois, Iowa, Michigan, Missouri, Nebraska, Oklahoma, Tennessee, and Wisconsin.
To report this story—and others you’ll see on our website soon—we use an approach rooted in the scientific method that we are calling The Markup Method:
Build. We ask questions and collect or build the datasets we need to test our hypotheses.
Bulletproof. We bulletproof our stories through a unique review process, inviting external experts and even the subjects of investigations to challenge our findings.
Show our work. We share our research methods by publishing our datasets and our code. And we explain our approach in detailed methodological write-ups.
In the case of our Allstate investigation, when we found that that Allstate had been pushing an algorithm that sets prices based on how likely you are to shop around, the first thing we did was build a dataset.
Investigative Data Journalist Maddy Varner and Investigative Reporter Aaron Sankin reviewed tens of thousands of pages of insurance filings before they found a document that detailed how the algorithm would affect every single Allstate customer in the state of Maryland—all 93,000 of them. The table was 1,101 pages long.