Reading companion · Cervas Election Study 2026 · American Political Divides & Great Debates
In Week 1 you answered the eight questions Douglas Ahler and Gaurav Sood put to a national sample in 2015. Each one asks what share of a party belongs to a group everyone associates with that party. We got them wrong in the same direction their sample did — every group looks bigger than it is — and by less: we overshot the truth by 14 points on average, their 1,000 adults by 20. This page is the long version of the argument made in class: the article’s theory, its evidence, its vocabulary, and our own answers held up against it, with every number traced to its source.
The exact wording of every question is reproduced at the bottom of the page.
These controls drive every figure on the page. The second one needs a word of explanation: three people answered all eight questions with the identical number — one put 68 on every question, one put 65, and one put 0. Nobody thinks every group is exactly 68% of its party, so that is the fingerprint of someone clicking straight down the page without reading the questions. Survey researchers call it straightlining. The switch takes those three out of every figure at once.
Leaners are counted with the party they lean toward. The highlighted row is the filter you have selected; the table follows the controls above.
Read the article first if you can; this page will make more sense as a second pass than as a substitute. Then work down the page in order. It follows the article’s own logic — what the parties are, what people think they are, whether the gap matters — and at each step it puts our Week 1 answers beside the national sample’s, so that you are reading the study as a participant rather than a bystander.
Under every figure there is a short guide in three parts: how to read it, what to notice, and ask yourself. The third part is the one to bring to class.
The title is a borrowing. In 1922 Walter Lippmann opened Public Opinion with a chapter called “The World Outside and the Pictures in Our Heads,” and argued that people act not on the world as it is but on a simplified picture of it — he gave the word stereotype its modern meaning in that book. Ahler and Sood take the phrase and point it at the two parties: whatever a Democrat or a Republican is, most of us carry a picture of one, and the picture arrives before any evidence does.
Why should a political scientist care about the accuracy of that picture? Because of what partisanship is thought to be. The American Voter (Campbell, Converse, Miller and Stokes, 1960) described party identification as a psychological attachment, formed early and held for life, more like a religious affiliation than a considered judgment about policy. Green, Palmquist and Schickler (Partisan Hearts and Minds, 2002) sharpened the account: people decide which party is theirs by asking what kinds of people belong to each and which assemblage of groups looks most like them. On this view a party is, in the mind of an ordinary citizen, a coalition of social groups before it is a set of positions.
If that is right, then beliefs about party composition are not trivia. They are the raw material of partisan identity, and they should shape how people feel about the other side. The article’s move is to notice that these beliefs can be measured — you can ask what share of Democrats are Black, or what share of Republicans earn more than $250,000 — and, because the true shares are also knowable from large surveys, they can be scored. Once beliefs can be scored, two further questions open: whether the errors have consequences, and whether correcting them changes anything. The whole article is those three questions in order.
The consequence Ahler and Sood care most about is affective polarization: not disagreement about policy, but dislike and distrust of the other party’s supporters as people. That term comes from work Sood coauthored (Iyengar, Sood and Lelkes 2012), which showed that partisans’ feelings toward the other party had cooled far more since the 1970s than their policy views had diverged. The standard instrument is the feeling thermometer, a 0–100 rating the ANES has asked since 1964; the gap between a person’s two party thermometers is the usual measure of affective polarization, and out-party warmth is the outcome in the experiments described below.
Before asking how badly we misjudge the parties, it is worth establishing what there is to misjudge. The group account of partisanship holds that people attach to a party because of which assemblage of groups it appears to contain — you look at a coalition, decide whether it looks like you, and join. That is a claim about composition, so it can be checked.
Ahler and Sood's uncomfortable observation is that the two coalitions are far less distinct than the account requires. Majorities of both parties are white, straight, religiously affiliated, and without a college degree. The groups that supposedly define each side are minorities inside it — sometimes small ones, and sometimes no larger than the share on the other side.
These are 2024 figures, computed from the ANES Cumulative Data File — the same file the Democracy’s Data ANES lab uses, so every number below is one you can reproduce yourself.
Figure 1
Where the coalitions differ, and where they barely do
Share of each party’s identifiers with each characteristic. Leaners are counted with the party they lean toward. Hover any row for the numbers.
Each row is one trait. The blue dot is the share of Democrats and Democratic leaners with that trait in 2024; the red dot is the share of Republicans and leaners. The grey line between them is the gap, in points, and the control above the chart sorts the rows either by that gap or in the order the traits are usually listed.
Read the dots as shares of a party, not shares of the country: 19% of Democrats are Black; this says nothing directly about what share of Black Americans are Democrats.
Read the top of the chart when it is sorted by difference. Republicans are barely more likely than Democrats to be 65 or older (23.3% against 25.1% — the gap runs the wrong way), and barely less likely to have someone in the household in a union (14.3% against 16.6%). Asian Americans are 3.7% of Republicans and 5.3% of Democrats. Three traits that do real work in the stereotype, and none of them separates the coalitions by more than two points.
Real differences do exist, and they are not small: race, education, and religion separate the coalitions clearly. But even there the majority story holds. The most Republican trait in the whole chart is being white — 76.7% of Republicans — and 53.2% of Democrats are white too.
The figures above are from 2024. Everything after this is scored against Ahler and Sood’s benchmarks from 2012 — because scoring against their benchmarks is what makes our answers comparable to their national sample. Where we could confirm that the 2012 ANES reproduces their number, the charts also carry a faint dashed tick for the same group in 2024. How each benchmark was built, what using it costs, and why the article’s South is not the Census South are set out in Where the numbers come from, at the end.
This page has two jobs. The first is to lay out the article's own evidence, which is what you are reading now. The second, everywhere after this, is to hold our answers up against theirs.
The design. In March 2015 the authors put eight questions to 1,000 American adults through YouGov. For each one, respondents typed a number between 0 and 100 into a box — no scale, no options. Four questions asked what share of Democrats belonged to a group associated with the Democratic Party; four asked the same of Republicans. The answers were compared against the groups' actual shares, estimated from the 2012 ANES and Pew's 2012 Religion & Public Life Project.
Three choices in that design are worth pausing on, because each closes off an easy objection. The open box means no response option could suggest an answer; a scale that ran 0–10–20 would have anchored people, and a scale that ran to 100 would have anchored them differently. Four items per party, balanced, means the result cannot be an artifact of one party’s coalition being more “visible” than the other’s. And an external benchmark means the truth is not hostage to the same sample that supplied the guesses. The cost of that last choice is that the truth is itself a survey estimate with error of its own — the confidence bands on every chart below are the honest way to carry that cost.
The headline number is not measured in percentage points but in multiples. Averaged across the eight groups, respondents put them at 4.4 times their real size — an overestimate of 342%, with a confidence interval running from 327% to 358%. The table below is that average taken apart.
The two groups that are genuinely large — Southern Republicans and evangelical Republicans, both above a third of their party — are the two the public gets nearly right. The group that is genuinely tiny is off by a factor of seventeen.
Notice what kind of number 342% is. It is an average of ratios, and ratios explode when the denominator is small: a guess of 38 for a truth of 2.2 contributes a 1,600% overestimate on its own. Reported in points instead, the national sample overshot by about 20 on average, which sounds far less dramatic and is the same data. Neither presentation is wrong; each answers a different question. The ratio asks “how many times too big is the picture?” and the points ask “how far off is the number?” This page uses points for comparisons across groups and keeps the article’s ratio for its headline, and tells you which is which.
These are not averages dragged upward by a careless minority. For all eight party-group pairs a majority of respondents overestimated the group's share, and for six of the eight more than 70% did. Medians sit only a few points below means, so outliers are not driving the result either.
This is the finding that turns a curiosity into a claim about the public. If the error were concentrated in a few inattentive respondents, the remedy would be better surveys. Because it is nearly everyone, the remedy — if there is one — has to be something that reaches nearly everyone, which is what the experiments at the end of the article try.
Everything from here on is comparison, and the article’s remaining findings appear where our data meets them — who is guessing, whether it is arithmetic, and whether it matters. Their sample of 1,000 adults is the benchmark; our 47 students are the case. Where we match them, the finding travels. Where we diverge — and we diverge on more than you would guess — the interesting question is whether that is because we are unusual, or because our sample is too small to say.
Ahler and Sood's central finding is that people wildly overestimate how much each party is made of its stereotypical groups. Their national sample overshot by an average of 342 percent. The chart below puts our answers next to theirs, with the true share marked in green.
Read each row of Figure 2 left to right: the solid green tick is the 2012 benchmark, the red line below it is how far past it we went, and the violet line above it is how far the 2015 national sample went. Both overshoots are numbered at the right. The faint dashed tick, where it appears, is the same group measured again in the 2024 ANES.
Figure 2
Every group looks bigger than it is — but ours look less big
Mean estimated share of each party belonging to the group, our respondents against the 2015 YouGov national sample.
Each row is one of the eight questions. The green tick is the 2012 benchmark the article used, sitting in its pale confidence band. The red dot is our class mean; the violet diamond is the 2015 national mean. The two numbers at the right are how many points past the benchmark each landed. Where a faint dashed tick appears, it is the same group in the 2024 ANES.
The chart is about direction and distance, not about the exact value of any one dot. Hover a row for the full numbers.
A mean hides how much we disagreed. Every dot below is one of you. Pick a question and watch the answers scatter — on some items the answers cluster tightly near the truth, and on others they run from single digits to eighty percent.
Figure 3
Where each of us landed
One dot per respondent. Drag your eye to the green line to see who was close.
Pick a question. Every dot is one respondent’s answer to it, placed on the 0–100 scale and stacked where answers pile up. The solid green line is the 2012 benchmark; the dashed line, where there is one, is 2024. Dots to the right of the solid line are over the benchmark. Tick the box to color the dots by the respondent’s party.
The bias is not symmetric. Republicans' estimates of Democratic composition are significantly more distorted than Democrats' own, and Democrats return the favor when describing Republicans.
The figure opens on their sample — 438 Democrats and 336 Republicans, leaners included, from Table 1 of the paper. Republicans put Black Americans at 46% of Democrats while Democrats themselves said 39%; Democrats put the share of Republicans earning over $250,000 at 44% while Republicans said 33%. On all eight items the out-party overshoots further.
Figure 4
The out-party is the one you get wrong
Mean estimate from Democrats and Democratic leaners against Republicans and Republican leaners. Leaners are grouped with the party they lean toward, as in the original study.
Each row is one group. The blue dot is the mean estimate given by Democrats and leaners; the red dot is the mean given by Republicans and leaners. The number at the right is the out-party estimate minus the in-party estimate — for a Democratic group, what Republicans said minus what Democrats said. Positive means the party doing the describing overshot further than the party being described.
The switch above the chart changes whose answers are plotted. It opens on Ahler and Sood’s 438 Democrats and 336 Republicans, because that is where the finding lives; our five Republicans are the second view, not the first.
We can run the same comparison on our own answers, with one enormous caveat printed on the figure itself: there are only five Republicans and Republican-leaners in the whole study.
Read the our respondents view the way you would read any cell with five people in it — as a hypothesis, not a result. One Republican changing their mind about evangelicals moves that row by several points. What the figure is genuinely good for is showing you the shape of the claim Ahler and Sood are making, so that you can say what evidence would settle it.
The obvious objection is that people are poor with percentages and small numbers get inflated by default. The authors tested it: a version of the survey required estimates to sum to 100 across mutually exclusive categories, which should discipline the guessing. For six of the eight groups the estimates barely moved. Telling people the actual population base rates did not fix it either. And independents, interestingly, are about as accurate as people describing their own side — which is hard to square with the idea that this is simply innumeracy.
The misperception is about the parties specifically, not about numbers in general.
Everything so far has been about the first half of Ahler and Sood's argument — that the misperception exists. The second half is the reason anyone cares: they argue the misperception does something, that people who imagine the other party as a collection of stereotypes end up disliking it more. Our survey asked the party thermometers too, but 47 respondents — five of them Republicans, most of them warm toward both parties — cannot test a claim like that, and this page does not pretend otherwise. What follows is their evidence.
This is the part that makes the paper matter. In their observational data, the size of a person's bias about out-party composition predicts both partisan social distance (how unhappy they would be to have the other side marry into the family, move next door, or work alongside them) and allegiance to their own party.
A correlation like that is suggestive and no more. People who dislike the other party might inflate its stereotypical groups because they dislike it, rather than the other way around; or some third thing — how much partisan media a person consumes, say — might drive both. Observational data cannot sort those stories out. What can is random assignment: if you decide by coin flip who gets the corrective information, then on average the groups differ in nothing else, and any difference in how they feel afterward has only one available cause.
So they ran the experiment. Respondents were split into three conditions: one simply told the true composition of the out-party, one asked to state their beliefs first, and a control group given neither.
How big is 6.4 points? The thermometer runs 0–100, and the typical partisan rates the other party somewhere in the cold twenties or thirties, so a shift of 6 or 7 points is a fifth or a quarter of the distance to neutral — from one short exposure to a few numbers. The third result is the strangest and, for a course like this one, the most useful: simply being made to state a belief as a number seems to make people notice that they have never checked it. That is a description of what the Week 1 survey did to you.
Set our eight means beside theirs and a pattern shows. We overshot the benchmark by about 14 points on average; their national sample overshot by about 20. Ours is the smaller error on six of the eight items, with the largest differences on the two groups they inflated most — Republicans earning over $250,000, where they said 38% and we said 23%, and union-household Democrats, where they said 39% and we said 26%. We were worse on two: atheist and agnostic Democrats (33% against their 29%, on a benchmark of 8.7%) and Southern Republicans (47% against their 40%, on a benchmark of 36%). On Republicans 65 and older the two samples are within a point of each other.
Several explanations are available, and our data cannot choose among them:
What our data can say is narrower and still worth saying: the misperception Ahler and Sood describe appeared, in the same direction, on all eight items, in a room full of people who study this for a living. It was smaller. It was not absent.
The quiz used to sit right here, which was self-defeating: by this point you have already been shown the answers three times over. It now has its own page, so it can be taken before anything is given away.
Send it to someone who has not read this. It asks the eight questions in a shuffled order — as Ahler and Sood did, and as our form could not — makes you commit to all eight before it reveals anything, and then scores you against the truth, against us, and against the 2015 national sample.
Every number on this page comes from one of the items below, reproduced exactly as it appeared in the survey you took. Wording is not a footnote in this literature — it is the finding. Ask "what percentage of Democrats are Black" and you get one answer; ask people to sort Democrats into Black, White, Latino and other so the shares must total 100, and Ahler and Sood show the estimates move. Judge the figures above against the sentences that produced them.
Answered by typing a number into a box. No slider, no scale, no response options — which is why one answer in our data is 0.5.
Asked on a 0–100 scale, where 0 is coldest and 100 warmest. Two of them are the party thermometers, the instrument behind the article’s measure of affective polarization; the others rate the eight groups themselves. None is used in a figure on this page; they are reproduced because they were part of the survey you took.
The item that sorts everyone into blocs in Figure 4, and colors the dots in Figure 3:
Respondents picked one option. Following Ahler and Sood, leaners are grouped with the party they lean toward, which gives the three blocs used throughout:
Ahler and Sood asked the same eight quantities, and we borrowed their phrasing closely — including their substitution of "earning more than $250,000 per year" for the vaguer "rich." Three differences are worth naming before you compare the columns:
The true shares come from the 2012 ANES and Pew 2012, which is what Ahler and Sood benchmarked against. You answered in August 2026. Both parties have changed since 2012 — the Black share of Democrats has fallen from 22% to 19%, the gay, lesbian and bisexual share has doubled, Republicans have gotten slightly older, and Democrats with no religious preference have risen from 28% to 35%. Some of our apparent "error" on items like atheist or agnostic Democrats is probably real drift in the underlying quantity, not a misperception. Treat the green tick as the 2012 truth, not today's.
Forty-seven respondents from two of Prof. Cervas's courses — 28 from American Political Divides and Great Debates and 19 from Democracy's Data — mostly undergraduates, 29 Democrats or Democratic leaners to 5 Republicans or Republican leaners. The filter at the top splits the two courses apart, but note what that does to the party cells: Political Divides contains exactly one Republican, so its party comparison cannot be computed at all. Ahler and Sood find that out-party estimates are the biased ones — and almost all of our estimates about Republicans are out-party estimates. A sample with the opposite composition would likely produce a different pattern of errors.
Two respondents entered the same number for every item (68 and 65), and one entered zero for all eight. That is the signature of someone clicking through. The switch at the top of the page removes them from every figure at once; the story does not change much when you do, which is itself worth knowing. This is what survey researchers call straightlining, and finding it is part of the job.
Nothing in this section changes a finding above; it is where each benchmark was built, so that any number on this page can be reproduced from the files in the Democracy’s Data lab folder.
Every comparison chart on this page carries a green mark labelled the true share, and it is worth being exact about what that is. It is not a count — and none of what follows is stated in the article itself. All of it lives in the online appendix, section OA 1.3, on a single page, partly in footnotes.
The short version: Ahler and Sood did not measure party composition in their own survey — you cannot ask 1,000 people what they think the parties look like and use the same 1,000 to establish what the parties do look like. They imported the benchmark from elsewhere:
The appendix publishes those intervals, so we can draw them. Every green tick on this page now sits in a pale green band — the article’s own 95% confidence interval for that benchmark. They run from 1.1 points wide for atheist and agnostic Democrats to 4.3 for Southern Republicans.
So the green line is a survey estimate, weighted, with sampling error of its own — and in one case it is not even a tabulation. The 8.7% for atheist and agnostic Democrats was never published by Pew; the authors reconstructed it with Bayes’s rule from other figures in the report. It is the softest number of the eight, and it is the one our class misses by the widest margin.
The Southern item is the best story here. Ahler and Sood do not use the Census South. They use the twelve states a majority of respondents called “Southern” in a 2014 FiveThirtyEight poll — a folk definition of a region rather than an administrative one. Our first attempt used the Census South and came in 3.8 points high.
The cumulative file carries state, so we could rebuild their definition and try again:
Their benchmark reproduces exactly — 35.7%, to the decimal — once the twelve states are the right twelve. What looked like our error was a difference between the Census Bureau’s South and the one people carry in their heads, which is a fitting thing to trip over in an article about the pictures in people’s heads.
And we can carry it forward, though not from the file we started in. The Cumulative File blanks state for 2024 — every geographic field below census region is empty for all 5,521 respondents. The 2024 Time Series release still carries it (V243001), with county and ZIP restricted. Rebuild the same twelve states there and the folk-South share of Republicans is 38.5% in 2024, against 35.7% in 2012. The Census South reads 41.8% in both files.
The figures above are from 2024. The benchmarks used to score our guesses in the rest of this page are Ahler and Sood's, from 2012 — because scoring against their benchmarks is what makes our answers comparable to their national sample. The parties moved in the twelve years between, and on a couple of items they moved a lot.
Four of the eight items appear in that table, and the reason is a rule we imposed on ourselves. For each item we recomputed the 2012 value from ANES and compared it to the benchmark the article used. Where our 2012 figure reproduced theirs, we trusted the 2024 figure and drew it. Where it did not, we left the item off the charts.
Three of the four come back exactly. The one that fails is not an error so much as a different question wearing the same name. Failing the check does not mean we are blind — the lower half of the table above shows what those measures did between 2012 and 2024, including the two we can now put numbers on. It means only that we will not draw them beside the article’s benchmark, where a reader would take them for the same quantity.
Exit polls are the usual place to look for region, and they were the first thing we reached for when the Southern item looked unrecoverable. They would have been the wrong tool: an exit poll describes people who voted, sorted by how they voted, while the article asks about people who identify with a party whether or not they turned out — the same mismatch that disqualifies the union measure here. The right answer was in ANES the whole time, in a file we had not opened.
Computed from the ANES Cumulative Data File 1948–2024 (release of February 5, 2026), the copy held in the Democracy’s Data ANES lab. 2024 pre-election respondents, n = 5,521, weighted by VCF0009z; party is VCF0301 coded 1–3 Democrat and 5–7 Republican, so leaners sit with the party they lean toward, as in the article. The variable behind each row is named in the table below.
What this file cannot tell us. Four of the article’s eight groups have no clean counterpart here, and rather than substitute a near-miss we left them out of the 2024 comparison entirely:
The first three of those are measurable in the 2024 Time Series file with some care, and an earlier draft of this page did exactly that. We removed them because a number you cannot reproduce from the file in your own lab folder is a number you should not have to take on trust.