Here’s a tale of three cities: Atlanta, New York and Detroit.

In all three cities, there is a high degree of racial segregation in the schools. White students go to schools with relatively few black and Hispanic students. Black and Hispanic students attend schools that don’t have many white students. When Sean Reardon, a sociologist at Stanford University, measures the racial isolation in a quantitative way, he finds that the schools in the three cities are “equally racially segregated.”

But the poverty rates in the schools are very different. In Atlanta, black students go to schools with very high poverty rates. The students in these schools tend to come from families whose income is low enough that the children qualify for free or reduced priced lunches, a federal measure of poverty. The white students in Atlanta tend to go to schools with very low poverty rates. In New York City, Reardon finds the same pattern, but not to the same extreme. Meanwhile, in Detroit, this pattern isn’t true at all. White and black students attend different schools, but the poverty levels are high in both white and black schools.

It turns out, according to Reardon’s calculations, that the differences in poverty rates between the black and white schools are very predictive of the achievement gaps between black and white students. “The achievement gap is very small, virtually zero in Detroit,” says Reardon. “It’s quite big, but not enormous, in New York City. And it’s among one of the two or three biggest in the country in Atlanta.”

This example arises from a new study of achievement gaps and racial segregation in nearly every school in the United States. In the study, Reardon finds that racial segregation is a very strong predictor of the gaps in academic achievement between white and black or Hispanic students, but it’s school poverty—not the student’s race—that accounts for these big gaps. When the difference in poverty rates between black and white schools is larger, the achievement gaps between black and white students are larger. When the difference in poverty rates between black and white schools is smaller, the achievement gaps are smaller. The two phenomena—racial segregation and economic inequality—are intertwined, because students of color are concentrated in high-poverty schools.

“There’s a common argument these days that maybe we should stop worrying about segregation and just create high-quality schools everywhere,” says Reardon. “This study shows that it doesn’t seem to be possible.”

Reardon says he couldn’t find a single school district in the country where black and Hispanic students were learning apart from white students and performing well with test scores that weren’t lagging behind those of white students. In the cases where achievement gaps were small, such as Detroit, achievement was low for both black and white students. They’re not models to copy.

“It doesn’t seem that we have any knowledge about how to create high-quality schools at scale under conditions of concentrated poverty,” says Reardon. “And if we can’t do that, then we have to do something about segregation. Otherwise, we’re consigning black and Hispanic and low-income students to schools that we don’t know how to make as good as other schools. The implication is that you have got to address segregation.”

The study, “Is Separate Still Unequal? New Evidence on School Segregation and Racial Academic Achievement Gaps,” is a draft paper, meaning that it hasn’t yet been published in a peer-reviewed journal and may still be revised. Reardon, along with four current and former colleagues at Stanford, analyzed 350 million test scores from 2009–2016, representing about 50 million students, as they attended public schools from third to eighth grades. To compare apples to apples, scores from different state tests were converted to a single national yardstick.

Achievement gaps between white students and students of color are large. On average, across the country, white students are scoring nearly two grade levels higher than black students—the difference between fifth and third-grade achievement, for example. But it varies a lot by school district. In some districts, the black-white gap was as small as one-third of a grade level. In other districts, the achievement gaps are three grade levels apart. The white-Hispanic achievement gap across the nation is smaller—between one and one-and-a-half grade levels.

It’s well known that high-income students perform better on tests than low-income students. Higher-income students tend to have better educated parents who not only may read and talk to their kids more, but also convey the importance of an education and set high academic expectations for their kids. What’s interesting in this study is that not only does the level of school segregation predict the size of the achievement gap between white and black students, it also predicts the rate at which the achievement gap grows as students progress from third to eighth grade.

In conjunction with this study, the Stanford Education Data Archive released its vast data trove on an interactive website, the “Opportunity Explorer,” where anyone can see the test scores for every public school in the United States. It allows you to compare test scores within a district or between schools in two different states. Parents and policymakers can compare schools in Wichita, KS, with those in Bangor, ME, for example. In a separate “learning rates” tab, it also shows test score patterns over time, so you can see how students improve as they progress from third to eighth grade—a measure of how much kids are learning during their school years. In a third “trends” tab, it shows how schools are performing over time, measuring student test scores in the same grade from year to year.

Reardon advises visitors to the website to avoid equating test scores with school quality. “The average test scores that kids have in schools or school districts are the results of all the opportunities these kids have had to learn their whole lives, at home, in the neighborhood, in preschool and in the school year,” Reardon says, “so it’s misleading to attribute average test scores solely to the school where they take the test.”

“If you want to know how good the schools are,” Reardon says, “a better but not perfect measure would be the learning rates, because those are measuring how fast kids are learning while they’re in school, regardless of where they started.”

This story was produced by The Hechinger Report, a nonprofit, independent news organization focused on inequality and innovation in education.