(Darron Cummings/AP)

Early voting is underway in many states, and things seem to look good for Democrats.

CNN suggests that Democrats have improved their standing in Arizona and Nevada, compared with 2012. Democrats also are voting at high rates in the key battlegrounds of Florida and North Carolina, according to news reports. The New York Times is estimating the November vote in North Carolina, using the early vote totals and other information. That model predicts a Hillary Clinton victory.

Since people’s actual votes are secret, these predictions are generally based on the number of registered Democrats and Republicans who have voted early. But how well does the early vote actually predict election outcomes?

Political scientist Seth Masket at FiveThirtyEight, who analyzed 2012 early vote data, argues that it isn’t a good predictor of the eventual outcome in a state. But using a different set of data, we find that early vote numbers tell us a lot about which candidate will win.

Data collected by the voter firm Catalist and the Atlas Project show the number of early and absentee ballots cast in the run-up to Election Day in 2008 and 2012 for every state where the early vote amounted to at least 10 percent of the eventual total vote.

By comparing the share of early votes cast by registered Democrats and Republicans to the final vote in those states, we can determine how well early voting predicted Barack Obama’s vote share when he ran against John McCain (in 2008) and Mitt Romney (in 2012). This calculation sets aside unaffiliated voters or those registered with a third party.

The chart below plots the percentage of early votes cast by registered Democrats in 2008 against Obama’s share of the two-party vote (setting aside minor party candidates). In the left-hand panel, we show this relationship for the early vote cast 10 days before Election Day.


The states are generally clustered around the diagonal reference line, indicating a reasonably strong relationship between the Democratic early vote and Obama’s vote share. (For the statistically inclined, the r-squared is 0.80.)

If a state falls below the line, its early vote overstated support for Obama. In North Carolina, for instance, 64 percent of the two-party early vote came from Democrats, but Obama received just over 50 percent once all the ballots were counted. On the other hand, in several states considered battlegrounds this year — Florida, Ohio, Nevada, and Arizona — the early vote was quite close to the eventual presidential vote.

Early voters have taken to Twitter and Instagram to share their experiences with long lines as they arrive at the polls to cast their votes. (Monica Akhtar/The Washington Post)

Notably, the right side of the figure, where we use the early vote just one day before Election Day, looks almost identical to the 10-day measure. (The r-squared is 0.82) In other words, the early vote 10 days before the election is just as good a predictor of the outcome as the final day’s early vote totals.

We find a similar pattern in 2012. By that year, several states had added early voting, giving us 19 altogether in the chart below.


Again, there is variation in the accuracy of the early vote 10 days out. Florida, North Carolina, Louisiana and Maryland overstated how well Obama would do, while Alaska, Idaho, Utah and California understated his support. In some of this year’s major battleground states, such as Colorado, Ohio, Nevada and Arizona, the early vote was a very good signal of the 2012 outcome.

Early voters have taken to Twitter and Instagram to share their experiences with long lines as they arrive at the polls to cast their votes. (Monica Akhtar/The Washington Post)

As one might expect, the early vote is more predictive of the eventual outcome in states where a higher share of the electorate votes early. For example, in 2012, the early vote explained 71 percent of the variance in the final vote in states where less than 25 percent of the vote came in early. But in states were more than 25 percent of the total vote was cast early, the early vote explained 83 percent of the variance.

In general, the party breakdown of the early vote — whether 10 days or one day out — tells us a decent amount about how that state will go.

Why do our conclusions differ from Masket’s? Our analysis of Catalist data for 2012 includes data from 19 states, compared to Masket’s 12. Additionally, Masket includes some states we excluded (such as Pennsylvania and Oklahoma) because early votes accounted for less than 10 percent of all votes cast. Indeed, early voting in Pennsylvania in 2012 was limited to absentee voting that required voters to provide an excuse.

None of this is to say that the early vote is more useful than other sources of information, such as polling or even the simple balance of party registration in a state. But in some battleground states, the early vote is closely related to the eventual outcome. Taken together with the polls, the numbers this year do indeed bode well for Clinton.

Brian Schaffner is professor of political science at the University of Massachusetts at Amherst. Find him on Twitter @b_schaffner.

Anthony Rentsch is a double major in political science and statistics at the University of Massachusetts at Amherst and a research assistant for UMass Poll. Find him on Twitter @Anthony_Rentsch.