PaperlayerView exact revision on arXiv

Source-linked AI summary

Understanding Human Perception of Representation in Citizens' Assemblies: An Empirical Study

Yusuf Hakan Kalayci, Vasilis Varsamis, Nick Gill, Evi Micha

arXiv:2609.27368v1cs.GTcs.CY

TL;DR

Quota-based assembly design requires choosing which attributes should define representation, even though panel sizes and recruitment constraints limit how many can be included. Using conjoint experiments and predictive models, the paper finds that party and climate concern each add predictive information, and their combination performs best.

Abstract

from arXiv · show

Citizens' assemblies are deliberative bodies intended to form a microcosm of the population. Organizers rely on quota-based stratification and must decide which attributes define resemblance to the public. Yet meeting every quota can still leave a dimension citizens value unrepresented. We study this attribute-selection problem in general-purpose and climate-focused assemblies through randomized conjoint experiments. We find that demographic attributes matter for perceived representation, but political alignment and context-specific attributes such as climate concern exert a stronger influence. When both are shown in a climate-focused setting, each remains influential, with political alignment having the larger estimated marginal effect. We also examine the omission of a relevant stratification attribute. Panels stratified on demographics, even with political alignment included, match the observed pool's climate-concern distribution no better than uniform random samples. These results suggest that representation on a relevant topic-specific attribute cannot always be recovered through correlated demographic or political quotas, and may therefore require explicit stratification. Finally, we ask whether representation preferences can be learned from the observed profiles. We compare predictive models, from simple, interpretable matching rules to a learned metric and a respondent-conditioned utility model. Both learned models predict choices for respondents excluded from training with substantial accuracy, revealing generalizable structure without fully capturing these judgments. Together, these findings guide attribute selection in citizens' assemblies: designers should consider political and topic-specific dimensions alongside demographics, avoid assuming correlated proxies protect omitted dimensions, and use predictive models to diagnose how profiles shape representation choices.

Introduction · Representation and Experimental Design

The paper asks which attributes people want represented and whether quota proxies can stand in for omitted dimensions; its conjoint findings show that political and topic-specific attitudes often weigh more than demographics.

Introduction

  • Because quota schemas cannot feasibly capture every intersectional profile, choosing which attributes to balance is unavoidable and should reflect which characteristics citizens value.Assemblies aim to make a small panel a microcosm of the wider public, commonly using self-reported gender, age, geography, and socioeconomic position.Source: Introduction · p. 1

Empirical Approach and Main Findings

  • Political alignment is the strongest representation cue in general-purpose and combined climate profiles, while climate concern dominates when political alignment is absent and outweighs demographic cues.Age-mismatch penalties are largest among respondents aged 16–29, and gender-mismatch penalties are larger among female than male respondents; these are descriptive comparisons, not formal tests.Source: Empirical Approach and Main Findings · p. 2
  • Adding party quotas to demographic quotas barely improves representation of omitted climate concern, while explicitly constraining climate concern substantially reduces representation error.The stress test applies LEXIMIN to the 1,151 Survey 1 respondents as the pool and compares uniform sampling with progressively expanded quota schemas.Source: Empirical Approach and Main Findings · p. 2
  • On held-out respondents, learned metric and respondent-conditioned Bradley–Terry models outperform matching heuristics, with only a modest accuracy gain from the more complex utility model.The models are proposed as tools for auditing panels and testing whether shared attribute-space proximity adequately describes representation judgments.Source: Empirical Approach and Main Findings · pp. 2–3

Related Work

  • Prior work examines political representation and mini-public legitimacy, but does not isolate which demographic, political, or topic-specific attributes make a prospective assembly member feel representative.The paper uses randomized candidate profiles to examine that question directly.Source: Related Work · p. 3

Representation and Experimental Design

  • The study addresses a gap in quota-based selection: methods operate on preselected attributes, while the relevant attributes for perceived representation remain unclear.It studies which candidate profile an individual believes better represents them through shared background, experiences, and concerns.Source: Introduction · p. 1, Representation and Experimental Design · p. 3

Survey: Topics and Samples

  • Two UK conjoint surveys used 1,151 and 576 respondents to compare candidate profiles with seven shared attributes, varying party alignment and climate concern across assembly contexts.Survey 1 yielded 5,755 pairwise choices per topic; Survey 2 yielded 5,760 choices with both political alignment and climate concern displayed.Source: Empirical Approach and Main Findings · p. 2, Survey: Topics and Samples · p. 3

Randomized Candidate Profiles and Survey Tasks

  • When party alignment and climate concern appear together, candidate pairs are sampled from a smoothed empirical joint distribution while balancing match and mismatch combinations.This preserves population-relevant party–concern combinations rather than sampling the two attributes independently.Source: Randomized Candidate Profiles and… · p. 4

Which Attributes Shape Perceived · Representation?

Political alignment and issue-specific attributes often weigh more than demographic attributes in perceived representation, though their relative importance varies by task. Exploratory subgroup patterns and open-ended responses further suggest that representation judgments can depend on respondents’ profiles and deliberative qualities, not just static resemblance.

Representation?

  • The analysis compares the average marginal effect of changing an attribute from a match to a mismatch, averaging over randomized states of other profile features.The same other-attribute mismatch realization is held in both counterfactual terms; independent randomization of mismatch events identifies the contrast. In the both-cues experiment, party and climate values are sampled jointly, but their mismatch indicators remain independent.Source: The Relative Importance of Candidate… · p. 4

The Relative Importance of Candidate Attributes

  • Political alignment has the largest estimated mismatch penalty in the general task, while climate concern leads in the concern-only Climate task; both remain the two largest cues when shown together.In the both-cues task, party mismatch has the greater estimated penalty; age and region generally form a second tier, while area type and disability status have smaller effects. Within-task contrasts are causal, but cross-task comparisons are descriptive.Source: The Relative Importance of Candidate… · p. 4, Figure 1 · p. 5

Exploratory Heterogeneity in Representation Judgments

  • The age-mismatch penalty is largest among respondents aged 16–29 in each task, and gender-mismatch penalties are larger among female than male respondents in each task.Age-mismatch estimates for respondents aged 16–29 are −.208 in the General task, −.168 in the concern-only Climate task, and −.121 in the both-cues task; in the General task, gender-mismatch estimates are −.124 for female and −.046 for male respondents. These are exploratory point estimates, not formal tests of group differences.Source: Exploratory Heterogeneity in… · p. 5, Figure 2 · p. 6
  • Age preferences show an ordinal gradient, while climate-concern preferences are directionally asymmetric: very concerned respondents favor their own value over lower-concern candidates, unlike fairly concerned respondents.The youngest respondents show the clearest decline as candidate age increases. Political-alignment estimates also show pairing-specific asymmetry; these patterns motivate graded gaps and reveal limits of a shared symmetric distance.Source: Exploratory Heterogeneity in… · p. 5
  • Open-ended feedback included one response emphasizing agreement over demographic background and another emphasizing how representatives reason and make decisions.The first response warned that demographic matching could encourage stereotyping; the second said static attitudes and opinions alone were insufficient.Source: Exploratory Heterogeneity in… · p. 5
  • Among 576 screened Survey 2 respondents, 179 (31.1%) named at least one additional theme beyond the displayed political-alignment or climate-concern cues.The most frequent suggestions were deliberative qualities (27 respondents), socioeconomic position (22), occupation or work experience (20), family or household status (19), local or community connection (19), capability or education (19), and climate knowledge or evidence orientation (17). Seven respondents also de-emphasized one or more displayed demographic attributes.Source: Exploratory Heterogeneity in… · pp. 5–6
  • The comments distinguish group-level inclusion, which demographic quotas can safeguard, from individual representation judgments that may also depend on substantive agreement and deliberative capacity.Because the optional free-text responses were self-selected, they identify considerations for future study rather than their prevalence or a ready-made set of additional quotas.Source: Exploratory Heterogeneity in… · p. 6

Achieved Indirectly?

Demographic and party quotas did not reliably preserve climate-concern representation, while directly constraining concern substantially reduced imbalance. The study tests whether correlated included attributes can stand in for an omitted attribute in quota-based selection.

Achieved Indirectly?

  • The design question is whether correlated included attributes can represent an omitted relevant attribute, avoiding the costs of collecting additional attributes.Otherwise, a panel can satisfy its stated quotas yet remain systematically unbalanced on an important omitted dimension.Source: Achieved Indirectly? · p. 6

Party Quotas as a Proxy for Climate Concern

  • In Survey 1, party affiliation and climate concern had a moderate association (Cramér’s V=0.261), making this a relatively favorable test of the proxy hypothesis.The study drew 250 panels of 40 from the 1,151-person observed respondent pool, comparing uniform sampling with demographic, demographic-plus-party, and direct concern quotas; the pool was the reference distribution, not the UK population.Source: Party Quotas as a Proxy for Climate… · p. 7
  • Mean TV was 0.0939 under uniform sampling, 0.0945 with demographic quotas, 0.0910 with added party quotas, and 0.0353 with direct concern quotas.Adding party reduced mean TV by only 0.0035 relative to demographic quotas; robustness checks found no consistent improvement from demographic quotas, with or without party, while direct concern constraints consistently improved balance over uniform sampling.Source: Figure 4 · p. 7, Party Quotas as a Proxy for Climate… · pp. 7–8

Predicting Representation Choices

The section evaluates how well alternative models predict individual representation choices on held-out respondents. Conditional Bradley–Terry performs best, while weighted ℓ1 offers a more parsimonious approximation; party and climate concern contribute complementary predictive information.

Predicting Representation Choices

  • Held-out choice prediction tests whether a single shared metric can summarize similarity judgments and can estimate which candidate better represents an observed respondent profile.Aggregated estimates could complement quota-based audits by identifying groups for whom available candidates are systematically less representative.Source: Predicting Representation Choices · p. 8

Predictive Models

  • The predictive models range from three interpretable heuristics to weighted ℓ1 proximity and conditional Bradley–Terry, which allows attribute importance to vary with respondents’ profiles.The heuristics use match counts, the largest training AMCE, or an AMCE-ordered matching tree; weighted ℓ1 permits unequal attribute weights and ordinal gaps, while conditional BT includes same-attribute respondent–candidate interactions.Source: Predictive Models · p. 8

Participant-Held-Out Evaluation

  • The evaluation uses 20 reproducible 85/15 respondent-level train–test splits, with exact prediction ties receiving half credit.Each respondent contributes five comparisons, so comparison-level and respondent-averaged accuracy coincide; task-environment accuracy differences are descriptive because task remits, displayed cue sets, and Survey 2 samples differ.Source: Participant-Held-Out Evaluation · p. 8

What the Comparisons Reveal

  • Conditional BT attains the highest held-out accuracy in every task, while weighted ℓ1 outperforms equal match count and the AMCE-ordered tree in every task.Conditional BT’s advantage is modest relative to its greater complexity, and its larger train–test gap is consistent with some overfitting. In the concern-only Climate task, weighted ℓ1 predicts at least four of five choices correctly for 64.7% of held-out respondents, compared with 67.5% for conditional BT.Source: Table 1 · p. 9, What the Comparisons Reveal · p. 9
  • Weighted ℓ1 held-out accuracy is .655 with party, .646 with concern, and .687 with both; conditional BT accuracy is .670, .658, and .710, respectively.The higher accuracy with both cues is consistent with party and concern each supplying predictive information not contained in the other, holding the seven demographic attributes fixed.Source: What the Comparisons Reveal · p. 9

Limitations and Future Work

The findings are bounded by the experimental tasks, profiles, and samples, and predictive accuracy does not establish democratic legitimacy or welfare. Future work should test broader contexts and attributes, connect individual judgments to panel evaluations, and develop constrained selection procedures.

Limitations

  • The experiments identify effects within displayed profiles and tested configurations, not their generalization to practice or complete-panel legitimacy.The conjoint tasks compared synthetic individual profiles; composition charts could require more demanding multi-dimensional trade-offs. Cross-task comparisons are descriptive, subgroup and value-specific patterns exploratory, and the two UK samples limit generalizability; quota findings are qualitatively stable across tested configurations, but exact magnitudes are pool- and configuration-specific.Source: Limitations · p. 9
  • Predictive accuracy reflects generalization within survey environments, not democratic legitimacy or welfare, and residual errors leave room for unobserved considerations and individual variation.Perceived representation is only one assembly objective; descriptive representativeness, substantive inclusion, deliberative quality, and public legitimacy can diverge (OECD 2020; Pow, van Dijk, and Marien 2020). Similarity optimization could reproduce homophily, so predictive scores should remain diagnostic and retain explicit inclusion safeguards.Source: Limitations · pp. 9–10

Future Work

  • Future work should replicate across countries, policy domains, and assembly remits, test more attributes, and investigate variation in how populations weight characteristics.It should also study how individual judgments aggregate into evaluations of complete panels before and after deliberation, and whether experience changes those judgments. Selection procedures could optimize estimated perceived-representation models subject to explicit diversity, inclusion, and fairness constraints.Source: Future Work · p. 10

Further Related Work

Related work situates citizens’ assembly selection within deliberative democracy, sortition, and representational fairness, while connecting this study’s empirical preference data to metric-based selection and metric learning. The cited work also addresses practical panel selection and learning distances from preferences.

Citizens’ assemblies, sortition, and deliberative democracy

  • Deliberative-democracy theory grounds legitimacy in informed deliberation, while citizens’ assemblies have proliferated and gained documented public support.Source: Further Related Work · p. 13
  • Selection research addresses representativeness, transparency, and intersectional representation (Baharav and Flanigan 2024), as well as dropout robustness (Assos et al. 2025; Gambhir, Flanigan, and Roth 2025).Permanent citizens’ assemblies heighten the importance of principled selection, and Sana, Wu, and Wells (2026) connect sortition to AI alignment through representative preference data.Source: Further Related Work · p. 13

Metric-based proportional representation and fair selection

  • Metric-distortion and fair-selection research studies committee representation and proportional fairness; this work contributes empirical preference data to construct and validate the metric spaces such algorithms assume.Related methods include the Expanding Approvals Rule and FairGreedyCapture (Ebadian and Micha 2025).Source: Further Related Work · p. 13

Metric learning from preferences

  • Metric learning infers distances from data, and this study evaluates it using direct pairwise choices and repeated participant-level train–test splits that keep respondents partitioned.Prior work includes Mahalanobis metric learning and learning from limited pairwise preference comparisons (Xing et al. 2002; Weinberger and Saul 2009; Wang, So, and Vinayak 2024).Source: Further Related Work · p. 13

Disclosure of AI Usage

The authors used AI for literature discovery, prose refinement, and analysis-code assistance, while retaining responsibility for the paper’s content and intellectual contributions.

Literature discovery

  • AI helped identify and summarize potentially relevant prior work, but authors independently checked relevance and accuracy before adding any citations.No citation was included without author verification.Source: Disclosure of AI Usage · p. 13

Prose rewriting and polishing

  • AI refined author-written drafts for clarity, readability, consistency, and grammar; authors reviewed the resulting text for accuracy and faithfulness.Source: Disclosure of AI Usage · p. 13

Analysis code

  • AI assisted with writing and debugging analysis scripts, while authors made all analysis-logic and experimental-design decisions and reviewed the code.Examples included AMCE estimation, metric-learning pipelines, and plotting.Source: Disclosure of AI Usage · p. 13

Scope limitations

  • AI did not generate core ideas, research questions, arguments, designs, formal or mathematical content, tables, figures, or structural changes; authors claim full intellectual responsibility.The paper identifies Claude Sonnet 4.5 and Opus 4.6 through Claude Code as the tools used.Source: Disclosure of AI Usage · p. 13

Experiment Details

The experiments used two UK surveys with screened samples, participant-facing assembly contexts, and randomized candidate profiles. The methods also describe how responses were cleaned and linked for reproducible analysis.

Fielding, Screening, and Analysis Samples

  • Survey 1 retained 1,151 respondents and Survey 2 retained 576 after screening; the surveys used different pairwise-choice and ranking designs.Survey 1 ran in four waves of approximately 300 and presented five pairs per task environment plus rankings; Survey 2 ran in two waves and presented ten pairs. Screening applied 2nd–98th percentile completion-time cutoffs after initial validity exclusions.Source: Fielding, Screening, and Analysis… · pp. 13–14

Participant-Facing Citizens’ Assembly Context

  • Participants received written assembly and task context, and each pairwise screen repeated that they should judge who would better represent them.The surveys made an explainer video available but did not require or record viewing, and included no assembly-comprehension check.Source: Participant-Facing Citizens’ Assembly… · pp. 15–16

Candidate-Profile Randomization

  • Candidate attributes generally matched respondents independently with probability 1/2, while Survey 2 sampled political alignment and climate concern jointly using a pre-screening party-by-concern matrix.For Survey 2, independently drawn match indicators gave each of four match–mismatch states probability 1/4; values within each state retained the smoothed empirical association.Source: Candidate-Profile Randomization · p. 16

Cleaning and Reproducibility

  • The cleaning pipeline standardized attributes, validated task and condition data, preserved within-survey linkage through pseudonyms, and stored analysis inputs and outputs for reproducibility.Participants provided informed consent through Prolific.Source: Cleaning and Reproducibility · p. 16

AMCE Analysis Details

The analysis estimates mismatch effects by averaging within eligible respondents before averaging across respondents, then uses exploratory subgroup and value-specific analyses to examine variation beyond aggregate effects.

Estimator and Uncertainty

  • The mismatch AMCE averages each eligible respondent’s mismatch-versus-match selection-rate difference, weighting respondents equally; a negative estimate means mismatching candidates are selected less often.A respondent is eligible only if they encountered both match and mismatch states.Source: Estimator and Uncertainty · p. 17
  • Uncertainty uses the standard error s_a/√N_a and a 95% Student-t interval; eligible respondent counts can vary by attribute because both match and mismatch states are required.Source: Estimator and Uncertainty · p. 17

Exploratory Subgroup and Value-Specific Analyses

  • Filtered AMCEs reapply the respondent-first estimator within subgroups, require at least 30 respondents and eligible respondents, and are descriptive rather than formal tests of subgroup differences.Other filtered analyses lacked recurring patterns clear enough to discuss; their omission is not evidence that no other subgroup heterogeneity exists.Source: Exploratory Subgroup and… · pp. 17–18
  • Value-specific heatmaps compare a candidate matching respondents’ own value with a specified alternative; zero denotes an even split, negative values favor the own-value candidate, and cells below 10 respondents are omitted.Source: Exploratory Subgroup and… · p. 18
  • Political-party mismatch effects vary by pairing and direction, so binary mismatch contrasts and a single weighted-ℓ1 weight cannot distinguish all respondent–candidate pairings.The political-alignment heatmap shows that particular party comparisons are not interchangeable.Source: Exploratory Subgroup and… · pp. 18–19, Figure 6 · p. 18

Open-Ended Attribute Elicitation and Survey Feedback

Survey 1 solicited general feedback, while Survey 2 asked for attributes missing from the displayed profiles. Because the prompts differed and Survey 2 responses were limited to nominations, the findings are illustrative rather than evidence for quota choices.

Survey 1 feedback

  • Survey 1 comments raised concerns about demographic stereotyping and static profiles’ inability to capture how representatives reason and decide.Because Survey 1 requested general feedback rather than additional attributes, these comments were illustrative and excluded from Survey 2 coding and counts.Source: Open-Ended Attribute Elicitation and… · p. 19

Survey 2 question and sample

Survey 2 coding

  • Coding assigned one or more themes per non-empty response, with 82 responses receiving multiple labels and repeated political-alignment or climate-concern answers excluded from additional themes.“No additional attribute” was coded separately, though seven respondents also emphasized at least one attribute.Source: Open-Ended Attribute Elicitation and… · p. 19

Interpretation and limitations

  • Survey 2 themes are nominations for future study, not evidence that each attribute belongs in a quota schema.The item was answered by less than half of the cleaned sample, and mention counts do not measure the strength or priority of preferences; the two surveys are not directly comparable.Source: Open-Ended Attribute Elicitation and… · pp. 19–20

Omitted-Attribute Analysis Details

In the observed Survey 1 pool, demographic and party quotas did not reliably reproduce climate-concern composition, while directly constraining concern reduced panel error. This pattern persisted across the tested panel sizes and quota tolerances.

Observed Pool and Party–Concern Association

  • In the 1,151-person Survey 1 pool, party and concern had an association of V = 0.261 (p < 10^-16), but this aggregate association did not ensure concern balance in feasible panels.The reference distribution was the observed pool, not the UK population.Source: Observed Pool and Party–Concern… · p. 20

Primary LEXIMIN Experiment

Panel-Size, Tolerance, and Quota Robustness

  • Across k=20, 40, 60, and 80 and exact to ±2-seat tolerances, directly constrained concern beat uniform in every feasible grid cell, while demographic and party quotas showed no consistent advantage.The auxiliary simulations used random linear objectives, not LEXIMIN; their boxplots summarize variation across generated panels rather than population confidence intervals.Source: Panel-Size, Tolerance, and Quota… · p. 21

Predictive Models and Results Across Tasks

Across three task environments, conditional Bradley–Terry has the highest mean held-out accuracy, though its larger train–test gaps qualify that result. In the both-cues task, party and climate concern each add predictive information beyond the other, and participant-level results show conditional Bradley–Terry improves prediction without radically changing who it fits well.

Common Evaluation Design

  • All models are evaluated on 20 shared 85/15 participant-level splits, with no respondent in both partitions and outcome-based preprocessing restricted to training data.The General and Climate Assemblies each contain 5,755 comparisons from 1,151 participants; the both-cues task contains 5,760 comparisons from 576 participants. Exact ties receive half credit, and split dispersion is not a population-level confidence interval.Source: Common Evaluation Design · p. 21

Formal Model Specifications

  • The heuristic orders attributes from training estimates, without test outcomes; the average Climate Assembly order is climate concern, region, age, education, ethnicity, gender, area type, and disability status.In the both-cues task, the heuristic selects party; substituting climate concern yields training accuracy .575 and test accuracy .572.Source: Formal Model Specifications · pp. 22–23

Results Across All Task Environments

  • Conditional Bradley–Terry has the highest mean held-out accuracy in every task, while weighted ℓ2 never improves on weighted ℓ1.The mean held-out accuracy of weighted ℓ2 is lower than weighted ℓ1 by 0.1 percentage points in the General Assembly, 2.2 points in the Climate Assembly, and 0.5 points in the both-cues task. Conditional Bradley–Terry’s training-to-test gaps are 3.4, 2.6, and 4.0 points, versus 0.7, 0.2, and 0.5 for weighted ℓ1; the larger gaps are consistent with possible overfitting but do not establish its cause. Split-level standard deviations are descriptive, not inferential; ℓ1 is retained because it is no less accurate and its additive distance is easier to interpret, not because it is generally superior.Source: Table 10 · p. 24, Figure 8 · p. 24, Results Across All Task Environments · pp. 24–25

Complementary Predictive Information from Party and Climate Concern

  • With both cues included, test accuracy is .687 for weighted ℓ1 and .710 for conditional Bradley–Terry, exceeding either cue alone in each model.Weighted ℓ1 accuracy is .655 with party and .646 with concern; conditional Bradley–Terry is .670 with party and .658 with concern. All four paired accuracy and log-loss intervals exclude zero, indicating that each cue contributes predictive information after the other within this experiment, under both model families.Source: Table 11 · p. 25, Complementary Predictive Information… · pp. 25–26, Table 12 · p. 25

Participant-Level Predictive Consistency

  • In the both-cues task, conditional Bradley–Terry predicts at least eight of ten choices correctly for 42.2% of test respondents, versus 34.8% for weighted ℓ1.Perfect prediction remains uncommon (5.5% versus 3.8%), and the two models’ respondent-level consistency distributions overlap substantially; conditional Bradley–Terry improves prediction without producing a qualitatively different fit distribution.Source: Participant-Level Predictive Consistency · p. 26

Learned Weighted-ℓ1 Profiles

Learned weighted-ℓ1 importance highlights the task’s dominant cue, but its ranking can diverge from AMCE when attributes use different encodings; recoding them consistently removes the reversal.

Learned Weighted-ℓ1 Profiles

  • Party has the largest learned weight in the General Assembly, climate concern in the Climate Assembly, and concern and party the two largest in the both-cues task.The remaining weight is distributed across demographic attributes rather than collapsing onto the dominant cue.Source: Learned Weighted-ℓ1 Profiles · pp. 26–27, Table 13 · p. 27
  • In the both-cues task, AMCE ranks party above concern (.278 versus .218), while weighted-ℓ1 reverses them (.191 versus .340).The profiles agree on the dominant cue in the General Assembly and Climate Assembly, but weighted-ℓ1 also assigns relatively more importance to ordinal age and education than AMCE does.Source: Table 13 · p. 27, Learned Weighted-ℓ1 Profiles · p. 27
  • The reversal reflects different units: weighted-ℓ1 ordinal weights are per-unit-distance slopes, whereas AMCE averages any mismatch against a match.Concern’s penalty is concentrated in the far tail, so AMCE is diluted by weak near-miss draws while weighted-ℓ1 must use a large slope to reproduce that tail.Source: Learned Weighted-ℓ1 Profiles · p. 27
  • When concern, age, and education are recoded as binary match/mismatch variables, party again outweighs concern and age and education align with AMCE rankings.This supports encoding asymmetry, rather than substantive disagreement between the analyses, as the source of the reversal.Source: Learned Weighted-ℓ1 Profiles · p. 27
Loading 2609.27368v1…