The Great Sorting · Knowledge
The hope was always that a better-informed public would meet in the middle. In the American National Election Studies the opposite is true: the partisan gap on policy and race is widest among the people who know the most about politics — every year, and by a growing margin.
There is an old idea, older than survey research, that political disagreement is mostly a misunderstanding — that if citizens only knew more, the heat would come out of their quarrels and reasonable people would converge on reasonable answers. It is a flattering idea, and the American National Election Studies has been quietly refuting it for fifty years. The survey has long carried a handful of factual questions: which party controls the House, whether the respondent can name the Speaker, the vice president, the chief justice, whether they perceive a difference between the parties at all. Score each person by how many of these they get right and you have a rough but serviceable measure of political knowledge. Then ask a simple question of the data. Among Americans who know a lot about politics, is the gap between Democrats and Republicans on the issues smaller or larger than among those who know little?
It is larger. Much larger, and it always has been. On the seven-point scale that asks people to place themselves from liberal to conservative, the most knowledgeable Democrats and Republicans were already more than a full point further apart than the least knowledgeable in the 1970s. By the 2010s the knowledgeable were pulling apart at better than two points while the least informed barely moved. Knowledge does not pull partisans toward a common center. It walks them, in opposite directions, to the ends of the scale.
The better-informed are the more divided — at every reading
Partisan gap on liberal–conservative self-placement (weighted Republican mean minus Democratic mean, 7-point scale; higher = further apart), among high-knowledge vs. low-knowledge respondents, by year.
Take-away: more political knowledge goes with a wider partisan gap, not a narrower one — and the high-knowledge line climbs while the low-knowledge line stays flat.
The same pattern governs attitudes that are not about ideology in the abstract but about a concrete and freighted policy: whether the government should help Black Americans improve their position, or whether they should help themselves. Here too the partisan canyon is deepest among the informed, and here too it has widened — the gap‑gap on aid to Black Americans grew from +0.56 before 1990 to +1.07 in the 2010s, nearly doubling. On both the symbolic question and the racial one, the people best equipped to evaluate the evidence are the people whose answers are most predicted by their party.
The natural objection is that “political knowledge” is just a proxy for a college degree, and that what we are really watching is the diploma divide wearing a costume. So split the respondents twice over: compare high- and low-knowledge partisans within the college-educated, and again within those without a degree, and pool the two strata. If education were doing the work, the knowledge gap‑gap should shrink toward nothing once schooling is held fixed. It does not. Net of education, the lib–con gap‑gap still runs +1.19 before 1990 and +1.49 in the 2010s; on aid to Black Americans it more than doubles, +0.44 to +1.01. Two people with the same level of formal schooling sort more sharply by party when they know more about politics. It is the knowledge, not the credential.
How much wider the partisan gap runs among the well-informed
The gap‑gap: partisan gap among high-knowledge minus the gap among low-knowledge respondents, by year. The lighter line holds education fixed (computed within college and non-college, then pooled).
Take-away: holding education fixed barely dents the pattern — the knowledgeable sort harder whether or not they hold a degree.
The Enlightenment bet was that facts would dissolve our disagreements. Among the people who hold the most facts, the disagreements are the most complete.
What does it look like up close, in a single year? Take 2020, the largest and cleanest reading. Hand the liberal–conservative scale to low-knowledge Democrats and Republicans and their answers overlap heavily: the two party averages sit only about half a point apart, both clustered near the middle. Hand the same scale to high-knowledge partisans and the distributions march to opposite walls — informed Democrats average 2.7 on the seven-point scale, informed Republicans 5.6, a separation of nearly three points with little middle ground between them. The well-informed are not splitting the difference. They have sorted into two non-overlapping camps.
Up close, 2020: the informed camps barely overlap
Distribution of liberal–conservative self-placement (1 = most liberal, 7 = most conservative) by party, among low- vs. high-knowledge respondents. Weighted shares.
Take-away: among the least informed the two parties' curves sit nearly on top of each other; among the best informed they pull to opposite ends of the scale.
This is not what a knowledge deficit looks like. A deficit model predicts that information closes gaps; what the ANES shows is information opening them. The result lines up with a body of theory that long ago abandoned the deficit story. John Zaller's account of how Americans form opinions turns precisely on awareness: the politically attentive are the ones who reliably receive elite cues about what their side believes — and, crucially, the ones best able to resist the cues coming from the other side.1 Half a century ago Philip Converse had already located whatever ideological constraint exists in mass publics among the small, informed stratum that follows what goes with what;2 as that stratum has grown and the parties have sorted, its capacity to bind issues to party has become the engine of the divide rather than a check on it.
The sharper challenge to the optimistic story comes from the experimental work on motivated reasoning. Dan Kahan and colleagues gave people a data-interpretation problem and found that greater quantitative skill produced more political polarization, not less, once the problem was dressed as a partisan controversy: numeracy was recruited to defend the identity, not to find the answer.3 James Druckman and colleagues show the mechanism at the level of the environment — as elites polarize, citizens lean harder on party cues, discount substantive information, and grow more confident in the conclusions they reach that way.4 Knowledge, in this picture, is not a neutral solvent. It is a toolkit, and the well-equipped use it to build a sturdier wall.
None of this counsels ignorance, and none of it says the informed are irrational. It says that information is processed by people who already belong to a side, and that the better you understand the political world, the more cleanly you can fit each new fact into the story your side already tells. The flattering hope — that we disagree because we are underinformed — survives only where the data are thin. Where Americans know the most, they agree the least.
Data. ANES Cumulative File, weighted. The knowledge scale is the mean of a respondent's available factual items — knows which party holds the House, perceives a difference between the parties, can identify the Speaker, the vice president, and the chief justice — scored 0–1; respondents valid on at least two items are kept, with high-knowledge defined as a score of 0.67 or above and low-knowledge 0.34 or below. The partisan gap on an issue is the weighted Republican mean minus the weighted Democratic mean on the native seven-point item (leaners folded into Democrats and Republicans; higher values are more conservative). The “gap‑gap” is the high-knowledge gap minus the low-knowledge gap.
Net of education. To separate knowledge from schooling, the gap‑gap is recomputed within the college-educated and within those without a four-year degree, then pooled across the two strata by cell size. The pattern survives, so the sharpening is not merely the diploma divide; in several early years only one stratum carries enough partisans in both the high- and low-knowledge cells, which is why the education-net line is shorter than the raw line.
Honest limits. The single most important caveat is 2024, shown dashed throughout: only three of the five knowledge items were fielded that year (which party holds the House and the vice-president question were not asked), so the scale degenerates and the low-knowledge group swells to an implausible size — 871 low-knowledge against 185 high-knowledge, the reverse of every other recent year — and the 2024 gap‑gap collapses artificially to +0.49. We report the late-era averages as stated, with 2024 included, which is the conservative choice: it pulls the trend down. Drop the broken 2024 point and the post-2012 lib–con gap‑gap is +2.00, not +1.62. The finding is solid from 1972 through 2020. Separately, η-style measures aside, these are differences of weighted means, not causal effects: knowledge is observed, not assigned, and the informed differ from the uninformed in many correlated ways. The claim is descriptive and consistent across five decades, two very different issues, and the education-net cut.