Ideological compasses around a presidential election: the ideology-policy gap

Sebastián Freille
Instituto de Economía y Finanzas (UNC)

Collaborators: Rubio, M; Jaroszewski, V.; Hofmann, R.; Albarracin, C.; Prazoni, G.; Larovere, S.; Seia, L; Bachiglione, C.; Balbo, S.

2026-05-01

The Puzzle

A Familiar Paradox

Preferences about spending on health and education

A Familiar Paradox

90% of Argentines reject cuts to public health and education. Yet Milei won promising exactly that.

4 in 10 US Republicans support universal healthcare. Yet they vote for candidates who oppose it.

The question: Why do voters choose candidates whose policies they oppose?

The Standard Model Fails

Median Voter Theorem predicts convergence and policy-based voting.

Reality:

  • Increasing polarization
  • “Identity” voting
  • Apparent contradictions between beliefs and behavior

This paper: We measure the disconnect—and show it predicts voting.

Ideology vs. Policy: A Distinction

Ideology Policy
Nature Abstract, symbolic Concrete, operational
Example “Smaller government is better” “Privatize Aerolíneas Argentinas”
Literature Converse (1964); Ellis & Stimson (2012) Alesina et al. (2020)

Ellis & Stimson (2012): 25-32% of Americans are “conflicted conservatives”—identify conservative but support liberal policies.

Our contribution: Measure this spatially in two dimensions, link to vote choice.

Research Design

Two Parallel Questionnaires

Ideology Survey

  • 17 items (8 economic, 9 social)
  • Abstract, identitarian language
  • “Countries with smaller governments tend to be more successful”

Policy Survey

  • 17 parallel items
  • Concrete, operational language
  • “Privatize Aerolíneas, BNA, and ARSAT”, “Eliminate Tarjeta Alimentar and use those funds to reduce taxes on PyMEs”

Each respondent gets two positions on the political compass → the gap between them is our key variable.

Sample: Two Survey Waves

2024 Wave 2025 Wave Combined
N 25 23 48
Milei voters 15 19 34
Massa voters 4 3 7
Blank/No vote 6 1 7
  • FCE-UNC students surveyed around 2023 election
  • 2025 wave includes self-reported political identities (16 categories)
  • Pooled analysis with year controls

Candidate Positions via LLM

We locate candidates on the same compass using LLM-based inference:

  1. Fed Gemini 2.5 Pro transcripts of presidential debates
  2. Asked it to answer our two surveys as each candidate –we “asked” all 34 questions to each candidate
  3. Required verbatim quotes as justification

Result (ballotage candidates):

Candidate Economic Social
Milei +9.0 +3.0
Massa -6.75 -1.0

Main Results

The Political Compass: Where Voters Stand

Policy positions with candidate locations. Blue = Milei voters, Red = Massa voters.

The Gap: Ideology → Policy

Arrows show direction and magnitude of each voter’s ideology-policy gap.

The Representation Paradox

Voters are systematically more centrist than the candidates they vote for. Dashed arrows show distance to “own” candidate.

Inverse Median Voter?

Milei voters average 8.66 units from Milei. Massa voters average 3.04 units from Massa. Candidates are more extreme than their supporters!

Key Finding #1: The Gap Predicts Voting

Logistic regression: Distance to candidates in ideology and policy space

Model Accuracy N
Ideology distance only 58.5% 41
Policy distance only 48.8% 41
Logistic (ideology) 92.7% 41
Logistic (policy) 90.2% 41
Full model (ideology + policy) 92.7% 41

Robustness: Leave-one-out CV = 87.8%; Bootstrap 95% CI = [82.9%, 97.6%]

The Econometric Model

We transform the probabilistic voting model into a spatial loss function:

\[L^{ij} = \alpha \sqrt{(\sigma^{ij} - \sigma^A)^2} + \beta \sqrt{(q^{ij} - q^A)^2}\]

The logistic specification for vote probability:

\[P(V_i = \text{Milei}) = \frac{1}{1 + \exp\left( -\gamma_0 - \alpha \cdot \Delta d_{\text{ideo}}^{i} - \beta \cdot \Delta d_{\text{pol}}^{i} \right)}\]

Where \(\Delta d = d^{Massa} - d^{Milei}\) (positive = closer to Milei)

Predicted Probabilities: The S-Curve

Probability of voting Milei as a function of ideological and policy distance differentials. Steeper curves = stronger predictive power.

Model Validation: Confusion Matrix

Confusion matrix showing predicted vs. observed votes. 38 of 41 votes correctly classified (92.7% accuracy).

Key Insight

The model correctly identifies all 7 Massa voters and misclassifies only 3 of 34 Milei voters as Massa supporters.

Key Finding #2: Milei Voters Have Larger Gaps

Gap magnitude by vote choice. Milei voters show systematically larger ideology-policy divergence.

Quantifying the Gap Asymmetry

Vote Mean Gap Median Gap SD
Milei 3.21 2.60 1.77
Massa 2.06 2.44 1.32

Interpretation: Milei voters are more likely to be “conflicted”—their ideology and policy positions diverge more.

  • Connects to Ellis & Stimson’s “conflicted conservatism”
  • Extends finding to multidimensional space

Key Finding #3: Voter Typology

Four types of voters based on gap direction. “Conflicted” voters show ideology-policy mismatch.

The Four Types

Type Description % of Sample
Consistent Right Right ideology, right policy 39%
Consistent Left Left ideology, left policy 17%
Conflicted Right Right ideology, left policy 27%
Conflicted Left Left ideology, right policy 17%

~44% of voters show some form of ideological conflict—their policy stance doesn’t match their stated ideology.

Interpretation

What Have We Learned?

  1. Ideology ≠ Policy — 44% of voters show meaningful divergence between their ideological identity and policy preferences

  2. Asymmetric conflict — Right-wing voters exhibit larger gaps, consistent with “symbolic conservatism” (Ellis & Stimson)

  3. Predictive power — Spatial distances outperform simple proximity models by 35+ percentage points

  4. Inverse MVT — Candidates are more extreme than their voters, not the reverse

Connecting to the Literature

Finding Related Work
Ideology-policy gap exists Ellis & Stimson (2012); Claassen et al. (2015)
Gap varies by partisanship Achen & Bartels (2016); Mason (2018)
Voters more centrist than candidates Fiorina (2017); elite polarization
Policy misperceptions Alesina, Stantcheva & Teso (2018)

Our contribution: First spatial measurement of the gap at individual level, linked to vote choice.

Limitations & Next Steps

Limitations

  1. Sample: University students (FCE-UNC), not representative
  2. Single election: Argentina 2023 (high-stakes, polarized)
  3. LLM validation: Candidate positions need expert cross-validation

But: Proof of concept for methodology; replication underway.

Research Agenda

  1. Scale up: Online panels, cross-university, representative samples
  2. Identity mechanisms: 2025 wave has 16 identity categories—analysis in progress
  3. Panel design: Track same individuals across elections

Ultimate goal: Understand when and why ideology decouples from policy in voter decision-making.

Conclusion

Summary

  1. Methodology: Separate ideology from policy using parallel questionnaires; locate voters and candidates in 2D space

  2. Main finding: The ideology-policy gap predicts voting with 92.7% accuracy (robust to cross-validation)

  3. Substantive insight: ~44% of voters are “conflicted”—Milei voters especially show large gaps

  4. Implication: Voters may choose candidates based on identity despite disagreeing on policy

The Takeaway

Voters don’t vote their interests. They don’t even vote their stated beliefs. They vote the gap between who they think they are and what they actually want.

Thank you.

sfreille@unc.edu.ar

Appendix

A1: Full Regression Results

Coefficient estimates with 95% bootstrap confidence intervals.

A2: Bootstrap Accuracy Distribution

1,000 bootstrap replications. 95% CI: [82.9%, 97.6%].

A3: Compass by Survey Year

Policy positions faceted by 2024 vs 2025 cohort. Results stable across waves.

A4: Gap Distribution

Distribution of ideology-policy gap magnitude across all voters.

A5: Gap Direction Patterns

Scatterplot of economic vs. social gap direction. Quadrants show conflict types.

A6: Political Identity Distribution (2025 Wave)

Self-reported political identities. Multiple selections allowed.

A7: Axes by Political Identity

Ideology and policy positions grouped by self-reported identity.

A8: Voter Centroids vs. Candidates

Average position of Milei and Massa voters vs. candidate positions. Dashed lines show distance to “own” candidate.

A9: Model Accuracy Comparison

Model Correct Total Accuracy
Ideology-based (nearest candidate) 24 41 58.5%
Policy-based (nearest candidate) 20 41 48.8%
Combined (average distance) 22 41 53.7%
Logistic regression (ideology) 38 41 92.7%
Logistic regression (policy) 37 41 90.2%
Logistic regression (full) 38 41 92.7%

A10: Descriptive Statistics by Vote

Group N Econ Ideo Social Ideo Econ Policy Social Policy
Milei 34 +1.54 +0.22 +0.47 +1.53
Massa 7 -3.57 -3.47 -4.11 -2.52
Overall 41 +0.67 -0.41 -0.31 +0.84

A11: LLM Methodology

We used Google Gemini 2.5 Pro to locate candidates:

  1. Fed transcript of first presidential debate (Oct 1, 2023)
  2. Asked model to answer our survey as each candidate
  3. Required verbatim quotes as justification
  4. Cross-validated with CNE platforms and interviews

Advantages: Replicable, low-cost, transparent

Limitation: Requires validation against expert surveys

A12: Survey Instruments

Ideology questionnaire

Policy questionnaire

A13: Year Stability Check

Year Accuracy N
2024 100.0% 20
2025 100.0% 21
Pooled 92.7% 41

Model performs well in both cohorts separately.

A14: Alternative Distance Metrics

Metric Accuracy
Euclidean distance 92.7%
Manhattan distance 90.2%

Results robust to distance specification.