After months of campaigning and days of counting ballots, the major news media outlets have projected that former vice president Joe Biden has defeated President Trump.

However, Biden didn’t do as well as public polls projected. Some groups unexpectedly appeared to have shifted toward Trump, such as Latinos. In rushing to understand what happened, some have relied on the National Exit Poll (NEP) conducted by Edison Research to form narratives about what happened and why. But that data source appears to have significant flaws — which could skew those narratives’ conclusions.

Specifically, the NEP’s estimates of who voted — what percentage of voters fall into any given demographic group — appear to be wrong. This kind of problem has plagued the NEP in the past and, apparently, it is an issue again this year. If the NEP’s estimate of who voted is incorrect, then the vote margins — the percent by which each demographic group voted for each candidate — could be incorrect. That can distort our picture of how different groups voted. And if the numbers for how different groups voted Trump/Biden are wrong, they shouldn’t be used to try to explain what happened in this election.

The NEP’s estimates are quite different from prior research

This year the NEP suggests that just 65 percent of voters were White and 34 percent were White without a four-year college degree. These estimates are dramatically smaller than what other research has found during prior elections. For example, the States of Change project — a series of reports that I co-authored with Ruy Teixeira and Bill Frey — found that 74 percent of voters were White in 2016, and 44 percent were White non-college. These estimates are identical to the Pew Research Center’s analysis of a large voter-validated survey.

What can that tell us about this year’s voters? We know that the relative turnout of different groups does not typically change dramatically between elections. If the relative turnout rates of different groups stayed the same, long-term demographic trends would lead us to expect 72 percent of 2020 voters to be White and 41 percent to be Whites without a college degree. For the NEP’s estimates of 65 and 34 percent to be correct, the relative turnout rates of different racial groups would have to have changed substantially and in ways that are not believable.

Here’s how I checked the data

To show this, I calculated the turnout rates that are implied by the NEP’s racial composition using several pieces of additional information. I started with estimates on the number of eligible voters in the country from the United States Election Project and combined them with projections about the racial composition of eligible voters from the States of Change project. This gives us a rough estimate for the raw number of eligible voters that fall into different racial groups.

From there, I took the projected final vote counts from the United States Election Project and combined them with the voter composition estimates that have been publicly released by the NEP. This gives us a rough estimate for the raw number of voters that fall into different racial groups.

Divide the raw number of voters in each group by the raw number of eligible voters in each group and you can derive the turnout rate implied by the NEP.

I find that the NEP implies that 66 percent of White Americans turned out in the last election. This is just barely higher than the implied turnout rates of Hispanic Americans (63 percent) and notably lower than the implied turnout rate of Americans who are Asian or belong to another racial and ethnic group (74 percent). That’s out of line with the U.S. Census Bureau’s Current Population Survey, widely considered to be one of the best sources of information about the U.S. electorate. The CPS has consistently shown that White citizens cast ballots at rates higher than those groups.

Further, analysis by political scientists Steven Ansolabehere, Bernard Fraga and Brian Schaffner suggests that the Current Population Survey is actually overstating the rate at which non-White voters cast ballots. This makes the NEP’s implied turnout rates — and the vote composition they are derived from — all the more unbelievable.

That means analysts trying to understand who voted how should not rely on the NEP

Let me be clear about my point of view. I am agnostic about the substance of the arguments people are making using the NEP. Maybe a given group did shift toward Trump or Biden and maybe it happened for the reasons being put forward by commentators. But the data source many people are using to make these arguments appears to have flaws that should give them pause.

Although the NEP survey will eventually be recalibrated based on the final election results — a process that may correct some of these issues — the current problems explained here suggest that analysts who wish to understand the 2020 election should find another source of data.

Robert Griffin is a political scientist and research director of the Democracy Fund Voter Study Group.