The only figures that are actual counts are certified results. The figures for white evangelical Protestants, and for everyone else, are estimates, not counts. Election results tell us how a state voted as a whole, but not how any group within it voted. We estimate each group’s vote from a large national survey of voters.
Every estimate made from a survey carries a margin of error (MoE). We calculate it in the standard way, known as the Wald method: we take the estimate’s standard error, a measure of how much the estimate would vary from one sample to another, and multiply it by 1.96. The estimate plus or minus that amount gives a range we can be 95 percent confident contains the true figure. The margin of error is half the width of that range. It reflects only sampling error, the uncertainty that comes from surveying some voters rather than all of them, and not other kinds of error (such as whether there are differences in the composition of those who answer surveys and those who don’t). The margins of error shown with the popular vote are rounded up to one decimal place, so that they are never understated.
In the electoral-vote panel of the National Popular & Electoral Vote visualization, our conservative estimate of change counts only definitive flips: states where not only the estimate but also the end of the 95 percent range least favorable to the estimated winner shows the state changing hands. (When either end of the range shows a different outcome from the estimate, we call the state “indeterminate.” It is treated as neither a definitive “flip” nor a “non-flip,” and its electoral votes stay with the winner of the certified vote.)
For the state-by-state results, we also checked two other ways of building the range. Both are “bootstrap” methods: they draw thousands of new samples from the survey’s own respondents, recalculate each estimate every time, and take the range from the results. One, the percentile method, uses the middle 95 percent of those results. The other, known as BCa, adjusts that range for any bias in them, a tendency to sit off-center from the estimate, and for any lopsidedness. The methods agree on nearly everything in the state results. Over five presidential election cycles, only one state that the Wald method counts as changing hands was indeterminate under another method: South Carolina in 2012, under BCa, in one of two runs. Overwhelmingly, the Wald method gave the more cautious end of the range.