When the economy loses legitimacy

An unfavorable view of the economy and disappointment with democratic performance can appear in the same interview. How closely do those judgments line up, and does the relationship look the same among people who are deeply dissatisfied and those who are relatively content?

In the pinned Round 11 European Social Survey, equal differences in economic satisfaction correspond to unequal differences in satisfaction with democracy. The fitted democratic-satisfaction difference is 1.99 points between economic ratings of 2 and 5, compared with 1.56 between 5 and 8. That is a bend in the association, not an identified point at which democracy loses legitimacy.

The comparison uses 47,836 paired responses across 30 countries in ESS Round 11, edition 4.2, conventionally labeled 2023/24. The whole source release spans March 11, 2023, to December 1, 2024. These are survey-period judgments, not a reading of public opinion in 2026.

The useful question is whether a straight line conceals something consequential, and whether the bend survives more demanding comparisons. We begin with the answers people actually gave, then ask what a model adds, which assumptions it needs, and where it stops being informative. Keep one country and round in view while examining those assumptions. France in Round 11 provides a concrete example; the headline remains the registered multi-country result.

Start with the answers, before trusting the curve

The two questions share a scale, but not a subject. One asks about the present state of the country's economy; the other asks about the way democracy works. Both run from 0, "Extremely dissatisfied," to 10, "Extremely satisfied." Dissatisfaction with democratic performance is not necessarily rejection of democracy as a regime, and neither response is an objective measure of institutional quality.

The mean compresses those answers into one number. The distribution layer restores the eleven actual categories, using the same paired-response population for both questions. The paired-response matrix asks a different question: which answers occur together? Its joint view shows a pair's share of all paired respondents; its conditional view changes the denominator to respondents at one economic score. Those percentages are not interchangeable, and a confidence interval around a mean is not the spread of people's answers.

Across the 30 included Round 11 countries, 31.0% of paired respondents give democratic performance a rating at or below 3 under the original ESS analysis weights (95% interval: 30.1% to 31.9%). That is a declared part of the response distribution, not a diagnosis of opposition to democracy. The weighted share and the unweighted count answer different questions.

For France, the observed, survey-weighted democratic-satisfaction mean at economic score 2 is 3.17, with a 95% interval from 2.85 to 3.49, based on 213 paired responses. At score 5, it is 5.10, with an interval from 4.78 to 5.42, based on 284 responses. At score 8 there are only 35 responses, so the public observed mean is suppressed. The gap marks insufficient support for displaying that mean; it is not a zero.

The distributions show how much variation a single curve must summarize. Now the model has to earn its place: does a simple straight line describe the association adequately, or does its convenience hide a meaningful difference?

A straight-line model makes a stronger claim than a general positive association. It requires every additional economic-rating point to correspond to the same fitted difference in democratic satisfaction. Two equal three-score intervals must therefore yield equal differences. Its upper-minus-lower contrast is exactly zero by construction, not because the survey has discovered zero curvature with perfect certainty.

The bend remains with country-specific starting levels

The first challenge is geography. A pooled curve can reflect country composition as well as associations within countries. The pooled spline gives an upper-minus-lower contrast of -0.46 points, with a 95% interval from -0.58 to -0.33. Allowing each country its own intercept retains a negative contrast. This model permits different starting levels, but assumes a common slope and curvature across countries; it does not estimate a separate national relationship.

The prespecified comparison makes the curvature readable. In the country-adjusted model, the fitted difference from economic scores 2 to 5 is 1.99 democratic-satisfaction points, with a 95% interval from 1.92 to 2.06. From 5 to 8, it is 1.56 points, with an interval from 1.46 to 1.66. Upper minus lower is -0.43 points, with an interval from -0.55 to -0.30. Each comparison spans three rating scores, not three years or an improvement in one person's circumstances.

For France, the fitted democratic-satisfaction levels at economic scores 2, 5 and 8 are 3.18, 5.17 and 6.73. The last number is a model prediction despite the suppressed observed mean. Selecting France retains its intercept, not a France-only gradient. Fitted levels standardize the remaining nuisance predictors at their analysis-weighted means.

The flexible model uses a smooth restricted cubic spline, with knots fixed at 0, 3, 7 and 10 before examining results. The joint test of its two nonlinear terms also rejects the straight-line restriction in Round 11 (p = 3.31e-11). But the knots are modeling choices, not discovered political fault lines. The smooth model tests neither a sharp threshold's location nor its nonexistence.

Adjustment must keep the respondents fixed

Country-specific starting levels address one source of difference, but people in the same country can face very different circumstances. Household financial difficulty is also a different question from satisfaction with the national economy. The further-adjusted model adds financial-difficulty categories, age and age squared, the source's binary gender coding, recorded education years and self-reported unemployment as main activity. That last measure is not an ILO unemployment rate.

Adding these fields reduces the Round 11 sample from 47,836 to 46,091 respondents. A fair comparison therefore first refits the country spline on those same complete cases. Its contrast is -0.43 points, with a 95% interval from -0.55 to -0.30. With the additional covariates, the contrast is -0.38, with an interval from -0.51 to -0.25. The negative equal-step contrast remains under this specification.

The two estimates come from overlapping fitted models, not independent experiments. Their separate intervals do not test the difference between them, and the change cannot be read as a percentage of the relationship "explained" by household circumstances. Matching respondents removes one ambiguity, sample composition, while leaving the interpretation of adjustment unresolved.

A probability asks what the mean cannot

The average curve answers one question well, but it cannot tell us how much probability sits among the most dissatisfied responses. A lower mean can arise through different changes in the answer distribution. The exploratory probability layer estimates the chance of a democratic-performance rating at or below each cutoff from 0 to 9. It uses ten separately fitted, survey-weighted binary models on the same 46,091 complete Round 11 records, with the original country, economic-score and additional adjustment terms.

The selected reference population matters. Fitted probabilities average predictions over its actual weighted covariate distribution; they are not predictions for an invented person with average characteristics. The uncertainty conditions on that fixed empirical reference. Every displayed target must pass an ordering check across all cutoffs and all eleven economic scores. If separate models cross, the whole target CDF is unavailable rather than sorted into a more convenient picture. The cutoff does not define rejection of democracy.

For the complete-case all-country reference, the adjusted chance of a democratic-performance rating at or below 3 is 53.2% at economic score 2 (95% interval: 51.6% to 54.8%), 16.5% at score 5 (15.6% to 17.6%) and 7.0% at score 8 (6.1% to 7.9%). The first three-score comparison is therefore a 36.7-percentage-point decline; the second is a 9.6-point decline. Upper minus lower is positive here, 27.1 percentage points (24.7 to 29.4), because the second change is less negative. These are standardized associations, not movements observed in the same people.

The ordering gate passes for 20 of the 32 reference populations, including France, the all-country reference and the fixed recent-country reference. Twelve national reference surfaces are withheld because separate cutoff fits cross somewhere on the registered grid. That is a limitation of the fitted distribution, not an absence of survey answers. The observed distributions remain available; rejected model surfaces are not selectively repaired.

The common curve is a restriction, not a European law

The shared curve is useful precisely because it summarizes many interviews, but that convenience is also an assumption. The original country-adjusted curve permits different intercepts, while retaining a common slope and curvature. The exploratory interaction models relax that restriction explicitly. One estimates country-specific shapes within a joint Round 11 fit. The other estimates shapes by reported household financial difficulty on the same complete-case sample, with the remaining covariates retained.

Each profile carries the same registered two-step comparison. The formal joint tests concern the common-shape restrictions; the individual intervals remain pointwise. A country's position is not a significant-country ranking, and a profile's respondent support is not the full model sample. Differences in conditional associations do not establish that financial circumstances caused a change in democratic attitudes.

The country common-shape test gives F(87, 22565) = 6.31, with p below 0.001. The financial-strain test gives F(9, 22114) = 2.67, p = 0.0042. These exploratory tests reject their respective shared-shape restrictions, but they do not establish that curvature rises monotonically with hardship. The four financial-profile curvature estimates are -0.30, -0.47, -0.05 and -0.42 rating points, in the source's order from living comfortably to finding income very difficult. The last two intervals include zero. A more elaborate model exposes that uncertainty rather than turning it into a tidy social gradient.

Earlier rounds test recurrence, not a European trajectory

A pattern in one survey wave might be peculiar to that period. Earlier rounds let us ask whether the chosen contrast recurs, although their changing coverage limits the comparison. Across the eleven separately fitted rounds, every country-spline contrast has a negative point estimate. The uncertainty is not uniform: the intervals include zero in Rounds 1, 3 and 4. Nor is this contrast interchangeable with the joint test of the entire nonlinear function. In Rounds 1 and 4, that joint test rejects a straight line even though the interval for this particular contrast crosses zero. Counting vaguely defined "significant waves" would obscure which question was tested.

Coverage also changes, from seven eligible countries in some early rounds to 30 in Round 11. Only Finland and France have reviewed usable designs in all eleven rounds. Those two countries cannot stand in for a fixed Europe, and the successive surveys are neither an individual panel nor an annual path.

A narrower check holds the 19 countries shared by Rounds 9 to 11 fixed. Their contrasts are -0.33 points in Round 9 (95% interval: -0.47 to -0.19) and -0.29 in Round 10 (-0.45 to -0.13). In Round 11, the contrast is -0.42 (-0.58 to -0.27), based on 29,719 paired responses. This supports recurrence within that recent country set; it does not establish a historical trend in the strength of the association.

A surviving result still needs a boundary

The final checks ask how much depends on the way the comparison was assembled. Restricting Round 11 to face-to-face interviews leaves 45,991 paired responses. The contrast is -0.43 points, with a 95% interval from -0.55 to -0.30, close to the primary estimate. Together with the complete-case and fixed-country checks, this shows that the reported bend survives several declared restrictions. It does not show that interview mode, missingness or country coverage is irrelevant.

The observed evidence deserves the same discipline as the models. Of 2,398 raw-score aggregates across the study, 640 are suppressed below 50 responses and another 303 are flagged for low precision at 50 to 99. An attractive fitted curve must not turn sparse observations into apparently precise local knowledge. France's suppressed high-score means remain suppressed even when a smooth model passes through that part of the scale.

Both answers are ordinal ratings on bounded scales. Equal numerical steps need not represent equal differences in underlying satisfaction, while a ceiling of 10 could compress differences toward the satisfied end. A common-response style could also contribute to the association between the answers. These are possible explanations for part of the association or its shape, not mechanisms established by this analysis.

Country intercepts and additional covariates also leave reverse causality and omitted influences unresolved. Respondents may use the rating scale differently across settings; the study does not establish measurement equivalence. Design-aware confidence intervals quantify sampling uncertainty under the model and survey conventions, not these broader uncertainties.

What the association permits us to say

For an analyst interpreting democratic dissatisfaction, a constant economic gradient discards information in the primary comparison. The association is steeper over the prespecified low-to-middle interval than over the equally wide middle-to-high interval, and that contrast persists under the declared adjustments.

The practical implication is to preserve the shape and its uncertainty when describing the evidence. A policy claim about what improving the economy would do requires a design that identifies causality. A claim about where democratic legitimacy collapses requires both a different political measure and a test designed to locate such a break. These satisfaction answers establish neither.