Multidimensional Political Attitudes and Polarization Across 141 Countries

Introduction

• Prior work had harmonized surveys or fitted measurement models, but not both at the individual level worldwide.Earlier international estimates were often country-year or party-level, while individual multidimensional estimates were confined to one survey program.Source: Introduction · p. 2
• Disjoint questionnaires create a bridging problem; shared items link instruments, and a bridge can also be constructed when no shared item exists.Afrobarometer and European Social Survey respondents answer questionnaires designed independently of one another.Source: Introduction · pp. 2–3
• The study places 1.14 million people from 141 countries on four common dimensions using harmonized survey items and a measurement model.The dimensions are social liberalism, economic left–right, democratic commitment, and anti-elite sentiment.Source: Introduction · p. 3
• 39 of 81 countries are most divided on a dimension that a left–right scale would miss, with the most-divided dimension varying regionally.The median within-country correlation between social and economic positions is −0.015.Source: Introduction · pp. 3–4
• Most political disagreement lies within countries, accounting for 58 to 88 percent of variance depending on the dimension.Ideological constraint exceeds +0.1 in 19 of 83 countries, and the United States reaches +0.70.Source: Introduction · p. 4
• The released item bank and fitted parameters allow future survey respondents to be placed on the same scale without re-estimation once items are harmonized.Individual estimates also permit comparisons between groups that country averages cannot provide.Source: Introduction · p. 5

Data, harmonization, and the common scale

• 1,145,632 respondents across 141 countries contribute 47,507,890 respondent-item observations to the assembled corpus.National election studies contribute 777 items, the largest single block.Source: Data, harmonization, and the common… · p. 5
• Of 1,299 harmonized items, 134 appear in multiple programs, and every pair of programs shares at least one item, connecting the corpus.Items are aligned manually from question text and codebooks, with independent validation and orientation checks.Source: Data, harmonization, and the common… · p. 5
• Four graded response models place respondents who answered at least two items on each dimension; marginal reliability ranges from 0.67 to 0.80.The 345 items not assigned to a dimension are excluded from all analyses.Source: Data, harmonization, and the common… · pp. 5–6
• Economic left–right has the least secure coverage: many countries rely on one redistribution item, Afrobarometer has no such attitude item, and verified links are few.Democratic commitment combines authoritarian values with satisfaction with democracy; anti-elite sentiment combines corruption perceptions with institutional distrust.Source: Data, harmonization, and the common… · p. 6
• Within-country correlations with left–right self-placement are +0.17 for social liberalism and +0.13 for economic left–right, while democratic commitment is +0.05 and anti-elite sentiment is −0.06.Agreement is stronger where the left–right frame is native and can be negative in post-communist countries.Source: Data, harmonization, and the common… · p. 6
• Cross-region item-discrimination correlations are 0.91 for anti-elite sentiment, 0.57 for economic left–right, and 0.28 for social liberalism; democratic commitment cannot be screened this way.The authors treat country-level social-liberalism quantities as an ordering rather than a measurement.Source: Data, harmonization, and the common… · p. 6

Most political disagreement lies within countries

• Country distributions are wide relative to distances between countries, and the four dimensions often place countries differently.Figure 1 shows distributions rather than country means; its ridge widths are uncorrected for measurement error.Source: Most political disagreement lies… · p. 7, Figure 1 · p. 8
• Country-mean estimators correlate at 0.99 for anti-elite sentiment and social liberalism, 0.96 for democratic commitment, and 0.70 for economic left–right.Joint country and program adjustments shift median positions by 0.17 standard deviations on social liberalism and 0.16 on democratic commitment; the authors report unadjusted positions because program effects combine multiple sources of variation.Source: Most political disagreement lies… · p. 7

The distributions behind the country means

• 88% of economic left–right variance, 75% of anti-elite variance, 76% of social-liberalism variance, and 58% of democratic-commitment variance lie within countries.These shares correct for measurement error and disagreement between survey programs. Countries with similar averages can nevertheless have markedly different internal spreads.Source: The distributions behind the country… · p. 9
• Country spreads correlate with item counts at −0.30 and survey-program counts at +0.45, so comparisons matched both sources of apparent disagreement.The matched country pairs differ by no more than a third in items answered and have the same number of covering programs.Source: The distributions behind the country… · p. 9

Demographic divisions

• 0.29 standard deviations separate over-65s from under-30s on social liberalism, the largest median demographic gap; demographic categories generally sort opinion weakly.Religious respondents are 0.17 standard deviations more conservative on social liberalism; top-third earners are 0.17 to the right on economic left–right; tertiary-educated respondents are 0.18 toward the democratic pole. The median residence gap is 0.07 to 0.12, while the United States has the largest urban–rural gaps on social liberalism and economic left–right.Source: Demographic divisions · p. 10
• For 21 of 24 division-by-dimension combinations, the middle 80% of countries spans zero, so gap directions vary across countries.Income on economic left–right is a near-universal exception: 91% of the 90 countries with measurable gaps put higher earners to the right.Source: Demographic divisions · p. 10
• The social-liberal gender gap grows younger: −0.020 among over-65s, −0.036 at ages 30–44, and −0.046 at ages 18–29.The youngest-cohort pattern reverses in 26 of 123 countries, where young women are more conservative than young men.Source: Demographic divisions · p. 11
• Single-wave country data cannot distinguish cohort differences from life-cycle differences or show whether today’s gender gaps persist as people age.Repeated waves scored on a common scale would be needed to distinguish the two explanations.Source: Demographic divisions · p. 11

Ideology coheres in a small set of countries

• Ideological constraint is a within-person relationship between positions, so country means cannot establish whether individuals organize opinions along a common line.The section tests where dimensions move together within individuals and among whom.Source: Ideology coheres in a small set of… · p. 11

Correlations are near zero in most countries and substantial in a few

• −0.015 is the median within-country social–economic correlation across 83 countries; 19 exceed +0.1, while Austria reaches −0.47.All six dimension-pair medians fall between −0.07 and +0.09. Correcting for measurement error moves the median to −0.021, and removing between-program differences moves it to −0.023.Source: Correlations are near zero in most… · pp. 13–14

Where there is coherence, it is concentrated among the attentive

• +0.12, +0.20, and +0.36 are the interest-tercile median correlations in coherent countries; all eleven put the most interested above the least.The gradient survives an out-of-sample split: the most interested exceed the least in 94% of country-splits among countries classified as coherent, versus 53% elsewhere. Item counts and posterior errors are similar across interest terciles, weakening the precision-artifact explanation.Source: Where there is coherence, it is… · pp. 15–16

Party-system polarization does not predict coherence

• Across 80 countries, party-system polarization correlates with mass coherence at +0.09 unweighted and −0.04 vote-weighted; neither estimate is distinguishable from zero.The largest observed association is +0.19, below the +0.31 minimum detectable at conventional power; this rules out an association as strong as the supply-side account implies, not every possible association. Switzerland has both a highly polarized party system and a coherent public (+0.35), unlike Hungary and Bangladesh despite similarly polarized party systems.Source: Party-system polarization does not… · pp. 16–17

A benchmark for claims of incoherence

• The coherence benchmark is a distribution, not a single number: the median country’s social–economic correlation is −0.015, versus +0.70 in the United States.An actor with a correlation of zero is typical by world standards, while the American case is unusual; the benchmark makes the choice of comparison explicit.Source: A benchmark for claims of incoherence · p. 17

Polarization is three separable quantities

• Spread is within-country standard deviation, bimodality tests for two peaks, and alignment measures whether positions across dimensions line up.Alignment restates the prior section’s coherence estimate; it is sorting, not spread.Source: Polarization is three separable… · p. 18
• Across 131 countries, spread and bimodality correlate at −0.15; across 83, bimodality and alignment correlate at −0.08, and spread and alignment at +0.28.The three measures therefore rank countries differently rather than capturing one underlying construct.Source: Polarization is three separable… · p. 18

Two corrections required for comparability

• The two corrections move mean country spread from 0.72 to 0.52; after program centering, North Macedonia ranks highest at 0.95.Austria’s error-corrected lead disappears after program centering; North Macedonia has only one program, so that correction cannot reach it.Source: Two corrections required for… · pp. 18–19
• Measurement error accounts for 35 to 40% of raw within-country variance; correcting it changes the spread–item-count correlation from −0.30 to −0.06.Respondents answer widely varying numbers of items, which changes the uncertainty in their estimated positions.Source: Two corrections required for… · pp. 18–19
• Before correction, measured spread rises with the number of survey programs covering a country at +0.45; centering estimates within country and source removes between-program differences.The correction leaves country positions untouched, but residual association with program count remains at +0.34.Source: Two corrections required for… · pp. 19–20
• The dip statistic cannot support a country ranking on bimodality because its values reflect item-score discreteness as well as two-peakedness.The point-estimate and posterior-draw versions correlate at only 0.08, and the authors treat the point-estimate country ordering as uninterpretable.Source: Two corrections required for… · p. 20
• Among 81 countries measured on all four dimensions, the most divisive dimension varies by region, including social liberalism in Europe and economic left–right in Latin America.Democratic commitment is most divisive in 67% of East Asian and Pacific countries and 50% of Middle Eastern and North African countries; only five Sub-Saharan African countries meet the four-dimension requirement.Source: What countries are divided about · p. 22

Measurement limitations

• These are cross-sectional estimates, not a time series; spread rankings also warrant caution because item discriminations vary across regions.The authors describe the corpus as a snapshot of the early 2020s and treat social-liberalism rankings as indicative.Source: Measurement limitations · pp. 22, 24

Discussion

• 58 to 88 percent of political-attitude variance lies within countries, and ideological constraint appears in only 19 of 83 testable countries.Demographic characteristics are weakly and inconsistently associated with attitudes; country means and corrected spread correlate between −0.24 and +0.22 across dimensions.Source: Discussion · p. 24

What individual-level estimates add

• Country means cannot reveal within-country variance, subgroup gaps, individual constraint, or how individuals rank against world opinion.For example, a model's percentile in world opinion requires knowing where people are, not only where countries are.Source: What individual-level estimates add · p. 25
• Demographic divisions are small, and ideological constraint is not a general feature of mass publics: it is detected in 19 of 83 countries and rises with political interest where present.Constraint is strongest in the United States, Canada, Germany, Finland, and Switzerland; the authors report that both parts of Converse's argument replicate.Source: What individual-level estimates add · p. 25

Measurement invariance is the central assumption

• Regional screening finds cross-region item-discrimination correlations of 0.91 for anti-elite sentiment, 0.57 for economic left–right, and 0.28 for social liberalism; democratic commitment cannot be screened this way.The authors treat social-liberalism country comparisons as an ordering, while describing anti-elite sentiment as supporting defensible cardinal comparisons.Source: Measurement invariance is the central… · p. 26
• The authors' diagnostics agree that anti-elite sentiment is the soundest dimension for cardinal comparisons and social liberalism the most fragile.Across instruments, offsets account for 7.8% of within-country variance on social liberalism versus 3.0% on democratic commitment; the corresponding regional-agreement values are 0.28 and −0.19.Source: Measurement invariance is the central… · pp. 26–27

Where the linkage is thinnest

• Only 134 of 1,299 items link the survey programs, and the election studies connect through 29 items, including just five economic items.The authors identify further verification of election-study economic items as the most valuable corpus extension.Source: Where the linkage is thinnest · p. 27

Other limits

• The scale's common-item assumption cannot be formally tested from within this corpus, and point-estimate uncertainty is not propagated to downstream comparisons.The authors use external diagnostics instead; on social liberalism, 198 of 219 items appear in only one program, leaving too few shared items for a formal test.Source: Other limits · p. 28
• The estimates are cross-sectional, but released item parameters allow future harmonized surveys to place respondents on the same scale without re-estimation.The authors propose a later pass through continuing survey programs to turn current cross-sectional comparisons into trends.Source: Other limits · p. 28

Harmonization checks and linkage

• Alignments group questions judged to measure the same construct, recode them to a common [0, 1] scale, and orient them semantically rather than by word similarity.The authors validate alignments through country-level agreement and agreement with other items in a dimension; a failed first check is repaired, while the second can reveal multidimensionality.Source: Harmonization checks and linkage · p. 29

Coding item direction

• Correcting the orientation of 277 source items changed estimates more than any other modeling choice in the pipeline.Mechanical direction procedures performed poorly: option-label rules were correct in 6 of 23 cases, and within-study correlations were invalid for this purpose.Source: Harmonization checks and linkage · p. 30, Coding item direction · p. 30

Harmonization checks and linkage

• 134 of 1,299 harmonized items link multiple programs, and every program pair shares at least one item, forming one connected corpus.The national election studies connect through 29 items, including five economic left–right items, which the authors identify as limiting that dimension's conclusions.Source: Harmonization checks and linkage · p. 30

Coding item direction

• Language-model-assisted direction coding agreed with hand coding on 52 of 56 items; only high-confidence answers were adopted, leaving the rest unpooled.A separate blind check agreed with hand coding on 16 of 18 disputed items; the per-source screen found sign disagreements in 32 of 1,386 cells, each treated as a candidate defect rather than a confirmed one.Source: Coding item direction · p. 31

Fitting the dimensions

• All four dimension fits converged, and fixed item parameters were used to score 1.14 million people on at least one dimension.Calibration samples ranged from 483,118 to 994,009 respondents, and countries needed 100 scored respondents for a country-level estimate.Source: Fitting the dimensions · p. 32

Dimension assignment and coverage

• Anti-elite sentiment covers all 141 countries, social liberalism 137, democratic commitment 140, and economic left–right 128.Items may belong to multiple dimensions, while 345 items remain unassigned.Source: Dimension assignment and coverage · p. 32

Respondent estimates and their standard errors

• Expected-a-posteriori scoring places respondents with at least two dimension items on a 61-point grid, and adding standard errors leaves point estimates unchanged.The stored social-liberalism θ values were bit-identical before and after standard errors were added; 1.14 million people have an estimate on at least one dimension.Source: Respondent estimates and their… · pp. 32–33

Polarization, and the corrections it requires

• The polarization analysis reports spread, Hartigan's dip statistic, and within-country alignment, using within-dimension percentiles to compare which dimension divides a country most.Countries need 400 scored respondents for spread comparisons, and spread and bimodality use respondents who answered at least three items; five countries are excluded from social-liberalism comparisons.Source: Polarization, and the corrections it… · pp. 33, 35
• 35 to 40% of raw within-country variance is estimation error, and correcting estimation error and survey-wave offsets moves mean spread from 0.72 to 0.52.The corrected ordering correlates +0.89 with the uncorrected ordering; wave-centering removes real change as well as instrument effects, which the authors cannot separate.Source: Polarization, and the corrections it… · pp. 33–34
• Bimodality rankings are not interpretable because the dip statistic tracks the discreteness of latent estimates, and plausible-value smoothing does not resolve the problem.The point-estimate and plausible-value dips correlate at 0.08, and the authors report no country ranking on bimodality as interpretable.Source: Polarization, and the corrections it… · p. 35

Constraint, and conditioning on it without circularity

• Out-of-sample classification avoids conditioning on the same tercile correlations used to assess the political-interest gradient.Among countries classified as coherent, median correlations rise from +0.13 to +0.21 to +0.36 across interest terciles; the most interested exceed the least in 94% of country-splits.Source: Constraint, and conditioning on it… · p. 36

Covariates

• The authors harmonize demographic and political variables from source releases, validate denomination coding, and keep self-placement and political interest outside the four fitted dimensions.Language-model denomination coding agreed with hand mappings on all 34 validation labels; the 18-source respondent-row alignment check returned rank correlations of 1.000.Source: Covariates · pp. 36, 38, 39
• Age, sex, and education reach 95–96% of respondents in all 141 countries, but income, internet use, and party attachment have narrower and nonrandom coverage.Income rank and economic strain are distinct measures and are reported separately; income-rank comparisons cover 94 countries and economic-strain comparisons 29, so neither supports a claim about the world.Source: Covariates · pp. 36–38

• The validation correlates latent estimates with self-placement within countries, avoiding a pooled measure that would mix individual and country ordering.Source: Validation against left–right… · p. 39
• Social-liberalism correlations with self-placement range from +0.51 in the United States to −0.20 in Slovakia.Correlations are also negative in Czechia, Latvia, and Bulgaria.Source: Validation against left–right… · p. 39
• Economic left–right correlates +0.47 with self-placement in the United States and +0.41 in Canada, but is approximately zero in Taiwan, Indonesia, Nigeria, and Tunisia.Source: Validation against left–right… · p. 41
• The authors read sign reversals in Slovakia, Czechia, Latvia and Bulgaria as predicted by party-system literature: there, “left” is associated with the former regime, inverting its relationship to social liberalism.They read this as a fact about the left–right frame rather than a defect in the scale.Source: Validation against left–right… · p. 41
• Attenuation makes the self-placement correlations lower bounds, because the single-item measure and latent estimates both carry error.Source: Validation against left–right… · p. 41

Screening for differential item functioning

• Regional fits use a shared item set to compare item discriminations across five blocks, making the screen interpretable without country-by-item estimation.Source: Screening for differential item… · p. 41
• Four of twenty regional fits reversed sign, so the authors anchor regional fits to the full-sample fit before comparing discriminations.Without sign correction, mirror-image fits would appear as large negative correlations and misleading evidence of item functioning problems.Source: Screening for differential item… · p. 41
• For MENA social liberalism, the −0.05 pooled correlation leaves regional item functioning undefined at this screen’s resolution.The authors therefore cannot say how social-liberalism items function in that block.Source: Screening for differential item… · p. 42
• The regional screen localizes where the invariance assumption strains but cannot establish invariance or correct for it.Source: Screening for differential item… · p. 42

Reproducibility

• A program-level metric model could not be estimated: 90% of social-liberalism items lack a second program, and no item spans all eleven program–waves as an anchor.The scalar model converged on shared items and placed program means 1.90 apart, but the metric model did not converge; the authors instead rely on three external diagnostics.Source: Reproducibility · p. 42
• The three country-spread specifications correlate with the uncorrected ranking at +0.94, +0.89, and +0.78, and substantive claims hold under all three.The paper reports the instrument-centered specification.Source: Reproducibility · p. 43

• The study uses secondary data from public, de-identified survey releases and recruited no human subjects.Source: Ethics statement · p. 43
• Individual estimates add no information beyond public responses and covariates, and are recoverable only by someone holding the source file with its pseudonymous identifiers.Source: Ethics statement · p. 43
• Distance from a median is a measurement, not a normative judgment, and estimates based on fewer items—concentrated in the global South—warrant greater caution.The authors note that additional survey items would contribute most in those less-covered places.Source: Ethics statement · p. 43

• The archive includes harmonization code, the alignment map, item bank, fitted parameters, individual estimates, and covariate code, organized as a pipeline from source files to results.Released fitted parameters allow a survey not yet seen to be placed on the scale without refitting.Source: Data and Code Availability · p. 44

The joint corpus

• Nine survey families contribute 1,299 distinct items and 47,507,890 responses from 1,145,632 respondents in 141 countries; national election studies contribute the largest block, with 777 items.The election studies supply a wide political-question battery with country coverage beyond the barometer regions.Source: The joint corpus · p. 50
• Alignments recode related items to a common [0, 1] scale and orient them semantically; 345 items that fail checks or fit no dimension are excluded from analysis.CSES modules are reported as a separate program for corpus composition but counted within the national election studies block for linkage analysis.Source: The joint corpus · p. 50, Table 2 · p. 52
• Every survey-program pair shares at least one item, creating a connected corpus; 29 of the 37 items linking national election studies to cross-national programs carry a fitted dimension.The remaining eight bridge items are unassigned; the 29 fitted items carry four dimensions, with several items loading on more than one.Source: The joint corpus · p. 50, Figure 7 · p. 51, Table 4 · p. 53
• Four graded-response dimensions are fitted on human data, with item parameters held fixed for scoring; economic left-right meets the within-country spread floor in only 93 countries.That floor requires 400 respondents per country who each answered at least three items for the dimension; the spread brackets omit countries with nonpositive corrected variance.Source: Table 3 · p. 52
• Figure 8 shows respondent distributions for 132 countries on social liberalism, while economic left-right covers 92; its ridges are not corrected for measurement error.Countries are ordered by mean social liberalism across panels, and fill indicates world region.Source: Figure 8 · p. 54
• Income rank and economic strain are separate, non-interchangeable measures, available in 94 and 29 countries respectively, with no respondents carrying both.Table 5 reports their subgroup gaps separately; no country meets its minimum group-size threshold for an economic-strain comparison on economic left-right.Source: Table 5 · p. 55