White evangelical Protestants versus everyone else

An examination of voting behavior from the last five presidential elections makes the vast difference vividly apparent. Yes, religion is a more important analytic lens than is generally appreciated.

Oct. 28, 2026 — With a week to the midterms, things are looking rosy for the Democrats, but gains may well be transient. There is no more profound weakness for Democrats than the segment of the electorate that is made up of white evangelical Protestants (20 percent or more). That’s not a startling new fact, true, but the skew is more consequential than is generally appreciated. Indeed, when one examines voting patterns in the last five presidential elections, it is as though white evangelical Protestants (WEPs) make up one country, and everyone else, another. 

Instead of two Trump victories sprinkled among Obama prevailing twice and Biden once (the certified results among all voters), the everyone else part of the electorate would have produced both popular and electoral-vote landslides for the Democrat every single time. 

The visualizations that follow detail how very different the everyone else part of the electorate is from WEPs at the national popular vote level. Then they primarily focus on how extensively the likely results of the everyone else part of the electorate would have shifted state-level results (and, accordingly, the award of electoral votes). The idea was to err on the side of caution when reporting apparent shifts.

The “what if” played out in this article raises important analytical and political questions. It is surely important to continue to assess the role of race, gender, age, education, and class in elections. (The New York Times just came out with a terrific set of visualizations covering all but the class element). But religion (including the specific religion-race match explored here) is another important lens that should not be ignored.

Politically, the WEP issue creates at least two problems that have not been solved. First, is there a way to cut into the nearly monolithic vote? (David French recently had a column exploring whether the race for United States Senator from Texas, with the Democratic candidate running what French calls “one of the most faith-forward campaigns in the country,” is apt to narrow the “God gap.”)

Second, if the idea of significant gains among WEP voters is illusory (or only available at the cost of significant losses among everyone else), then Democrats, it would appear, have less room to alienate remaining, potentially gettable non-WEP voters.

This article is based on both certified presidential election results in the 2008 through 2024 cycles and my analyses of data for those years from the Cooperative Election Study (CES),1 a survey instrument designed and led by researchers at Harvard and Tufts and fielded by YouGov. The CES describes itself as the “largest academic survey focused on American elections.” 

Prior to an election, YouGov will have gathered profile information from a respondent, including religion and born-again status.2  A survey wave in the weeks following the election asks respondents how they voted. 

All CES cycles but 2008 (a smaller sample) consisted of interviews with more than 50,000 adults and yielded between about 30,000 and 35,000 usable respondents (“usable” means the respondent provided his or her presidential vote and the CES validation process confirmed that the respondent did vote in that general election). 

For more detail on both the CES data and my analyses, please see the narrative accompanying each data visualization, notes in the data visualization, and the methods section at the end of this article. Last, I’d like to have your feedback: corrections, critiques, and suggestions, especially, though compliments are ok, too. Contact me here: craigurian at remappingdebate dot org.

For the purposes of these visualizations, white evangelical Protestants are white, non-Hispanic respondents who give their religion as Protestant or “something else,” and who say they are born-again or evangelical. (This is discussed further in a note in the visualization below.) 

In this first visualization, you see first the certified (actual) results of the election you select, both the popular and electoral vote split. When you select the “Everyone other than WEP” button (“Non-WEP only” in stripped down version), you get to see the estimates of popular vote among WEPs only and among everyone else. 

You also get to see how the electoral vote would change. Notice that we publish an “at least” number for Democrats, and, conversely, an “at most” number for Republicans. If one were to include states where the everyone else estimate showed a flip to the Democrat but where one end of the high-confidence range still showed a Republican result, the everyone else count would be even more lopsided. We don’t do this — instead calling those states “indeterminate” and retaining the certified result, as shown in the visualization on the next page — so as to keep the reported scope of change conservative. 

In the visualization below of the election at the state level, we start again with the certified vote. You see what happens when the “everyone other than WEP” button is selected. An estimate is provided for every state.3 Along with that estimate, we provide the end of the 95-percent range of reasonable statistical confidence that is least favorable to the estimated winner (“conservative” in a methods sense). We also show indeterminate states — those where the winner according to the conservative end of the range differs from the winner according to the estimate — as hatched and in the color of the certified winner.  (In the stripped down version, only states that flip or become indeterminate are listed.)

Flips from Republicans to Democrats are substantial, ranging from eight to 14 per cycle. Republican states that become indeterminate are common, too: they range from eight to 13, depending on cycle. (Across the five cycles, not a single Democratic state became a determinate Republican one, and only one Democratic state — New Hampshire in 2016 — became indeterminate.)

Two of the flips deserve brief mention. In 2008, only 32 percent of Mississippi respondents were confirmed as voters, whereas actual turnout was 61 percent. If the matching confirmed some kinds of voters more readily than others, the resulting sample could be skewed. That is unlikely to alter the flip status: the Democrat’s lead, even at the conservative end of the range, was large enough that only a severe skew could erase it.

Alaska in 2024 was another state where the estimate showed a Democratic win among everyone else. But the survey’s Alaska respondents as a whole reported voting for the Republican by about 33 points, while the certified result was only R+13. Matching the sample to the certified result therefore required an unusually large adjustment, and that adjustment assumes the survey’s error was shared evenly between WEPs and everyone else.4 If it was not, the result could be off by more than the (already large) range shows. Because the conservative end of the range was 2.3 points from a tie in these circumstances, this flip comes with a caution. 

Additional nuances are discussed in the methods section.

In this next visualization, the focus turns from flips to movement: regardless of whether a state flipped or became indeterminate, how much of a shift was there in the vote margin as between the certified results, on the one hand, and the everyone else hypothetical result on the other?

The starting point is the “Minimum” view. It shows the conservative end in a different sense from the previous visualization. There, the conservative end was the end least favorable to the estimated winner, because I wanted to be cautious about calling who would have won. Here, the conservative end is the end closest to no shift, because I wanted to be cautious about how much the result changes.

If you switch to “Estimated” view, you see the estimate itself, with its margin of error as the number after the ± sign. My own preference is to rely more on Minimum view.

In each view, the 25 states with the largest settled shift for the relevant dimension (minimum or estimate) are shown.5 

This last visualization delineates the WEP share of the electorate, both nationally and state-by-state. That share represents a substantial number of voters. Nationally, that is roughly 28 to 33 million voters who cast ballots for president, depending on cycle. And, there is a wide variation in WEP share among states: in several states and cycles about half of the voters are WEP. 

On wider screens, you can click or tap “Share” or “MoE” to sort on that basis. 

A key point regarding a state’s margin of error in respect to WEP share of the electorate. The margin of error obtained for the everyone else vote margin (page 3) and for the shift in margin shown as between the certified all voters result and the everyone else result (page 4) already take the margin of error for the WEP share into account. It would be a mistake to add any visualization’s margin of error to another’s. 

One further note on the margin of error on the WEP share. That number should be read together with the share itself, so that you understand the range of the WEP share. In Texas in 2008, for example, the WEP share was 32 percent, plus or minus 3. That makes it somewhere between 29 and 35 percent, so the share is well established. In Vermont in 2024, it was 16 percent, plus or minus 14: anywhere from 2 to 30 percent. That is a wide range relative to the share, so the WEP share figure is much less informative about how large the group is.

Methods

Much of the explanation of what I did is contained in the narrative sections introducing the article and preceding each visualization. More is contained in the text of the visualization itself and in hyperlinked notes in the visualizations.

This part of the article — which is not intended to be comprehensive — provides supplementation to what has come before. It is intended to be read after those other materials, as opposed to being a standalone document (indeed, this is not a standalone document from which one could fully understand the project). To the extent possible, material already discussed is not repeated here.

A. Sources of the data

1. The survey files we used came from Harvard Dataverse’s repository of CES files. We used each dataset’s latest version as of Oct. 27, 2026.

2. The certified election returns that we used, from the MIT Election Data and Science Lab, are in another Harvard Dataverse repository. In New York (in 2008, 2012, and 2016), the data as provided listed each ballot line a candidate appeared on separately, something that would have given the candidate’s non-Democratic-line and non-Republican-line votes to the “other” category. We added up each candidate’s votes across all the lines on which the candidate appeared.

B. Who is counted

3. We count respondents whose vote in that general election is confirmed by a voter file, who reported a vote for president for a Democrat, a Republican or another candidate, and who have a survey weight (see discussion on weighting in Section D, below). People who said they voted but whom no voter file confirms are not counted, and people the voter file could not find at all are left out rather than treated as nonvoters. There is one exception to this framework, described in Item 7. 

4. Here are the CES variables used to identify usable respondents, their presidential votes, and their weights:

5. The CES files did not contain a weight for 73 respondents in 2012 and 230 respondents in 2020 (0.24 percent and 0.64 percent of the cycle’s voter universe, respectively). We dropped them in lieu of imputing or borrowing another weight.

6. In 2012 in Oklahoma, three respondents reported a vote for someone other than the Democratic and Republican nominees, even though there was no one else on the ballot and no write-ins were permitted. As no such vote was possible, we dropped these respondents.

7. Uniquely across all states and cycles, Virginia in 2008 did not report vote-history records. According to the 2008 CES guide, “Virginia state law restricts use of vote history data, so all respondents from Virginia have a missing value.” What the CES file still did have was information on whether a respondent was actively registered or not. So what we did was to use “active registration” as the first criterion for inclusion (621 of 806 respondents were listed as active; 185 were inactive, dropped, unregistered, moved, or not found on the voter file). Naming a presidential candidate voted for was our second criterion, and 530 of the 621 did so. They were the respondents we counted.

8. I noted at the outset of the article that the survey universe was smaller in 2008 than in later cycles. Specifically, almost 33,000 people were interviewed, of whom almost 20,000 were usable respondents. Because of smaller numbers in 2008 (usable respondents varied between roughly 30 and 35 thousand in the other cycles), 2008’s ranges are correspondingly wider. Number of usable respondents who were white evangelical Protestants per cycle: 2008: 5,000; 2012: 6,521; 2016: 5,415; 2020: 5,875; 2024: 5,420.

C. Who is a white evangelical Protestant for purposes of this project

9. Directly prior to the first visualization, I explained that we treat respondents as white evangelical Protestants if they identify as white, non-Hispanics; give their religion as Protestant or “something else”; and say they are born-again or evangelical. As noted, anyone whose stated religion contradicts being an evangelical Protestant is out: Catholic, Mormon, Orthodox, Jewish, Muslim, Buddhist, Hindu, atheist, agnostic and “nothing in particular”. “Something else” is in, because answering it declines the list rather than naming a religion.

It is self-identification (as born-again or evangelical, and as “Protestant” or “something else”), not a respondent’s denomination, that we use. People in evangelical denominations who do not describe themselves as born-again or evangelical are not counted as WEPs.

10. The “something else” component of our WEP definition, as points of the electorate, in each cycle: 2008: 6.28; 2012: 0.98; 2016: 1.20; 2020: 1.94; 2024: 2.60.

11. In all cycles but 2008, there were separate race and Hispanic status questions, so both facts (white and non-Hispanic) are independently known. In 2008, there was one question which required a respondent to pick a single response from the races listed along with “Hispanic.” Where a respondent selected “white,” that respondent was treated as white, non-Hispanic.

We know from 2020 and 2024 that about two percent of white respondents were Hispanic and were excluded pursuant to our definition. If 2008 resembled those later cycles (and some white Hispanics were therefore included as WEPs), that would have represented an inflation of the 2008 WEP share by roughly 0.3 points.

D. Weighting, then calibration to the certified result

12. In 2012 through 2024, the survey provides a weight built after the election for respondents confirmed as voters, so that they resemble that year’s electorate. The 2008 file has no such weight. Its only weight makes the sample resemble U.S. citizens generally, not 2008’s voters. Our calibration corrects each state’s vote totals in every cycle, but in 2008 it cannot correct the makeup of the voters behind those totals, which in other cycles the survey’s own weight has already addressed. If, for example, 2008’s weighted voters had too many older or more educated people among Obama voters, calibration cannot fix that. As it happens, the 2008 WEP share in the survey’s own weighting (22.86 percent before our calibration) was almost a perfect match with the National Election Pool exit poll, conducted by Edison Research (22.81 percent), on comparable definitions.

13. Within each state we adjust those weights until the respondents’ weighted vote reproduces the state’s certified result for the Democrat, the Republican and everyone else, exactly. So our “actual” figures are the certified results, and the survey supplies only the WEP share and how each group voted. Nobody’s answers change, only how much each respondent counts.

14. Why calibrate? The other way of tying our figures to the certified result is to subtract the survey’s WEP vote from it. That would put all of the survey’s misreporting onto the result for everyone other than WEPs (the non-WEP result), worst in the close states where flips are decided. Calibration spreads the correction across all respondents.

15. What calibration assumes: It adjusts by reported vote alone, so it assumes the survey’s error is shared evenly between WEP and non-WEP respondents who reported the same vote. If white evangelicals misreported at a different rate, the non-WEP result would absorb a biased share, and no range covers it. The assumption matters most where the adjustment is largest. Calibration also has no test of fit: hitting the certified totals means the arithmetic closed, not that the weights are right.

16. In Oklahoma in 2008 and 2012, there was no certified third-party vote. In Rhode Island in 2008, there was a certified third-party vote, but no survey respondent there reported one. In both states, results were calibrated to the Democratic and Republican totals alone.

E. Estimates and their ranges

17. The non-WEP result is estimated directly from the non-WEP respondents in each calibrated state sample.

18. How the margin of error is computed. The estimate plus or minus 1.96 standard errors (the Wald method), with standard errors that carry the calibration through rather than treating the weights as fixed. Carrying the calibration through means this. A standard error measures how much an estimate would change if a different random set of people had been surveyed. Our weights are not simply the survey’s: in each state, we adjusted them until the respondents’ votes matched the certified result. Any other set of respondents would have been adjusted the same way, to the same certified result. So our standard errors are calculated as if each imagined sample had also been adjusted, which takes into account that every sample would be pinned to the same state totals. The alternative, treating our adjusted weights as if they had come with the survey, ignores that pinning and overstates the uncertainty. Taking the calibration into account gives smaller standard errors, and they are the accurate ones.

19. Every standard error was checked against a bootstrap (re-drawing the survey’s respondents at random thousands of times, redoing the calibration each time, and measuring how much the results vary), in every state and cycle. Our standard errors agreed almost exactly with the bootstrap’s in virtually every case. In the only three states where the two differed by more than 10 percent (Hawaii, 2008; Idaho, 2012; and Wyoming, 2024), using the larger figure changes nothing: Hawaii and Idaho stay no flips, and Wyoming stays indeterminate.

20. We examined two bootstrap methods (percentile and BCa), for every state in every cycle, and compared with Wald. The two bootstrap methods agree with Wald on nearly every call. Exceptions: South Carolina 2012, a flip under Wald, is indeterminate under BCa in one of two runs; New Hampshire 2016 is indeterminate under Wald and no flip under both others; and Nebraska 2012, Kansas 2008 and Nebraska 2008 change their call under one method in one run. Wald overwhelmingly gave the more cautious end.

21. The non-WEP range is computed from the non-WEP respondents, but a large non-WEP pool does not make it narrow. Calibration fixes each state’s totals, so chance variation among the WEP respondents carries straight into the non-WEP result. Its width depends mainly on how many WEP respondents there are and how large a share of the electorate they are. Alaska 2024 is an example: 14 WEP respondents at 35.5 percent of the electorate give a margin of error of 29.3 points.

22. A respondent’s weight decides how much they count, so a group of respondents can count for fewer than its number. Alaska 2024’s 14 WEP respondents count as 6.5 “effective respondents.” The project uses the Kish measure.

23. There are 11 cases, each a state in a given election year, where every WEP respondent voted the same way. In these cases, the sample shows no spread in the WEP vote, so the non-WEP range leaves out that source of uncertainty and understates it. It can matter only where the WEP share is large: Alaska 2016 (31.1 percent, 13 respondents; a flip, and present on both the Minimum and Estimated lists of margin shifts), Wyoming 2008 (26.5 percent, 7 respondents; indeterminate) and Wyoming 2024 (31.8 percent, 15 respondents; indeterminate). The other eight have shares of 8 percent or less. Outside the 11, the uncertainty in the WEP vote is included in the non-WEP range.

24. The ranges we report cover only sampling error. Other potential sources of error (for example, respondents misreporting their vote beyond what calibration corrects, answers carried forward from panel profiles that don’t match the respondent’s current views, who gets matched to a voter file, the make-up of the “Something else” box, respondents answering the race or religion questions in error, and the extent to which the calibration assumption is incorrect) are not covered. None of these has been quantified, and no width is invented for them.

25. Note two cases not already discussed in the article where the survey’s confirmed voters were unusually out of line with the state’s electorate. In Utah in 2024, 68 percent of respondents could not be found on a voter file at all (nationally, 34 percent). In Utah in 2020, of the respondents who were found, only 59.5 percent were confirmed as having voted (nationally, 81 percent). Neither affects a result: Utah is not a flip in either cycle, and is on neither of the Margin Shift lists.

F. National figures

26. National figures come from one estimate over all respondents after each state’s sample has been calibrated to its own certified result: never from a single national weighting, and never by adding up state estimates. This corrects for each state’s misreporting. In a test on 2024 data, it also had the smallest standard error of three approaches; the other two were versions of a single national weighting.

G. Cross-checks

27. Each cycle’s WEP share and vote were compared nationally with an independent survey: the exit polls (the National Election Pool exit polls, conducted by Edison Research, obtained from the Roper Center) for 2008, 2012 and 2016, and AP VoteCast for 2020 and 2024. Each comparison uses the closest match to our definition the other survey’s religion categories allow (the match is not identical), against our figures before calibration. These are only national figures, not state-by-state. The cross-checks feed no estimate: every figure in the visualizations comes from the CES.

28. So, two independent surveys, with different organizations, methods and ways of reaching people, find a group of roughly the same size, within about 5 points in each cycle, and almost exactly in 2008.

29. Now look at the margins among WEPs. Here, again, “Exit poll” refers to the National Election Pool exit poll, conducted by Edison Research, and obtained from the Roper Center.

30. This shows that WEP voters are very heavily Republican, both in our estimate and in the cross-check. That finding confirms the direction of the shifts, not the size of the shifts. As the table shows, our estimate, before calibration, of the national Republican lead among WEPs was very close to the cross-check only in 2012 and 2016. In the other cycles, our estimate of that lead among WEPs was roughly 9 to 13 points higher than the cross-check. The reasons for this were not investigated.

31. If the cross-check were closer to the truth, the shifts would be smaller than we estimated them, most of all in states where WEP voters represent a large share of all voters.

32. The cross-checks do not confirm that the definition is right, only that the measurement repeats. The other surveys ask essentially the same question, so agreement shows that independent surveys asking it get similar answers, not that the definition is right.

H. Miscellaneous

33. Note regarding any polling provided through the Roper Center. In no case were any respondent-level data seen or touched by any AI tool. When code written by Claude Code needed to be run on those data, I ran the code locally, never sharing any respondent-level data.

34. The rule about YouGov re-asking the religion and/or born-again questions where the stored answer is more than 14 months old is a rule that has been confirmed for the 2016, 2020, and 2024 cycles. It could not be confirmed for 2008 and 2012, for which no questionnaire was published. The guides for those earlier cycles say only that their religion and born-again questions were part of a profile survey taken before the main survey, without stating when stored answers were required to be re-asked.