Democrats need to win both of two Georgia runoffs to take control of the Senate. But despite President-elect Joe Biden carrying the state, they’re fighting an uphill battle in both races — particularly given they took fewer votes than Republicans the first time around in each race.

Biden won Georgia by about 12,000 votes, a very slim margin in a state where nearly 5 million were cast but enough to become the first Democratic presidential candidate to win the state since 1992. But now attention turns to the state’s runoffs. In those races, Biden received 100,000 more votes than Jon Ossoff, the Democratic candidate in the regular Senate election. Biden also outperformed the eight Democratic candidates in the special Senate election by about the same amount.

It is not uncommon for down-ballot candidates to receive fewer votes than the presidential candidates. But, of course, in the runoffs, in which Sen. David Perdue (R) will face Ossoff and Sen. Kelly Loeffler (R) will face the Rev. Raphael Warnock, there’s no presidential race. If Democrats want to win the Senate majority, they might need some of those who skipped the Senate lines on their ballots or voted for the Republican. And we can use precinct-level data to identify where this “undervoting” happened and give us an idea about who those people might be.

An average precinct in Georgia has just over 2,000 registered voters; these maps show the precinct breakdowns in the Democratic margin for president vs. the Democratic margin for Senate in Georgia. And then, we can zoom in on the Atlanta area (consisting of precincts within Cherokee, Clayton, Cobb, DeKalb, Douglas, Fayette, Fulton, Gwinnett, Henry, Rockdale counties).

Green areas are precincts in which Biden outperformed Ossoff, and orange areas are precincts in which Ossoff had a larger margin than Biden. The darker the color, the bigger the percentage point margin difference.

What do the data tell us?

Unsurprisingly, we can see that there are more green precincts than orange precincts: Biden did better than Ossoff all over the state. To get a better sense of who these voters are, we can look at data from L2 Political, a voter file vendor that collects data on registered voters across the United States. Using that data, we can examine the median household income, the fraction of voters with a bachelor’s degree or higher, and racial composition for each precinct.

If we look at precincts where Biden outperformed Ossoff the most (the 25 percent of precincts where he outperformed him by three points or more), we can get an indication or what types of voters supported Biden but not Ossoff. While this type of geographic analysis is not conclusive, it can point to a logical explanation.

In precincts statewide where Biden outperformed Ossoff, the median household income was $32,000 higher than the state average, the percentage with a bachelor’s degree or higher is 42 percent higher, and the population was slightly more White and less Black than the state overall.

Atlanta vs. not Atlanta

The picture is different in the counties that make up metro Atlanta and those outside it. Atlanta precincts are $30,000 wealthier, residents are 58 percent more likely to have a bachelor’s degree or higher. There’s also a difference in racial breakdown: Atlanta area precincts have a smaller proportion of White voters and more Black voters, on average, than outside Atlanta.

In the Atlanta area, precincts in which Biden outperformed Ossoff are $41,000 wealthier and have 37 percent more voters with at least a bachelor’s degree than the average precinct in those 10 counties. While in the average Atlanta area precinct 39 percent of voters are White and 43 percent are Black, in the Biden-but-not-Ossoff precincts, it’s 60 percent and 17 percent, respectively. What this tells us is that in Atlanta and its suburbs, it was people living in richer, Whiter and more educated precincts who voted for Biden and not Ossoff.

However, outside of Atlanta we see something else. The average household income and average education levels are nearly the same in precincts where Biden did and did not outperform Ossoff. Precincts in this part of the state are 65 percent White and 25 percent Black, but precincts in which Biden outperformed Ossoff are 53 percent White and 36 percent Black.

Obviously, not all voters in Georgia are White or Black. And there are differences among Hispanic and Asian American voters as well. However, since those groups make up only 5 percent and 2 percent of voters in an average precinct, respectively, the relative changes in those groups in average precincts compared with Biden-but-not-Ossoff precincts is relatively small.

What does it mean?

One possible explanation is that the type of voter who cast a ballot for Biden but not Ossoff is different inside and outside of Atlanta.

In the Atlanta area, it is possible that this kind of voter voted for Biden for president and Perdue (or Libertarian candidate Shane Hazel, who received nearly 115,000 votes). An indication is that in that part of the state only 18,000 people voted for president but not for Senate, but Ossoff trailed Biden by more than 70,000 votes. So this indicates that at least 52,000 people voted for Biden and a Senate candidate who was not Ossoff.

Outside of Atlanta, the undervote number is 28,000 — just shy of the 30,000 who voted for Biden and not Ossoff. So it seems more likely here that people who voted for Biden skipped the Senate race overall. (It’s harder to compare with the special, where Warnock got just 32.9 percent of the overall vote in a crowded field with multiple candidates from each party.)

What else might be going on?

Generally, drawing conclusions on individual voter choices based on aggregate geographic data (called ecological inference) is fraught.

You can read more about ecological inference and what can be done about it here. But the thing to keep in mind is that the average proportion in a precinct and the voters whose behavior we are trying to analyze might be different.

For example, outside of Atlanta we identified that in precincts with a higher percentage of Black voters, Biden outperformed Ossoff. But the precincts still have non-Black voters, and while it’s possible the Black voters caused the difference, it’s also possible they didn’t. It could be that White voters in majority Black precincts vote differently than White voters in majority-White precincts, and those voters are the ones who voted for Biden and not Ossoff.

It’s an election analysis trope to say that elections come down to turnout. But in this runoff, following the highest-turnout election in more than a century, the winners will be determined by which party can more effectively mobilize its base. However, there seem to be some voters who aren’t necessarily part of the Democratic base but are willing to vote for Democrats for president. A core question is whether Democrats can persuade these voters to help them gain a Senate majority — and whether their reluctance to vote for the full Democratic ticket the first time around will hurt them the second time.

This article originally misstated how many people must have voted for Biden and a Senate candidate who wasn’t Ossoff in the Atlanta area. It has been updated.

Anthony Pesce contributed to this report.