motief/docs/solutions/best-practices/overton-extended-analysis-methodology-2026-05-26.md

module	date	problem_type	tags
analysis/right_wing	2026-05-26	best_practice	[overton-window voting-margin mechanism-classification party-differentiation left-wing-response 2d-extremity]

Analytical Extensions for Overton Window Shift Analysis

Context

The core methodology in overton-window-shift-methodology-2026-05-24.md established the 7-step framework (strict centrist definition, Procrustes-aligned SVD, centrist support fraction, opposition control, extremity stratification, 2D audit). These are prerequisites — this document covers extensions built on that foundation for richer, more defensible analysis.

Domain decomposition (domain-decomposition-overton-analysis.md) must also be applied before interpretation to avoid conflating strategic moderation with genuine acceptance expansion.

Guidance

1. Replace Binary Pass Rate with Voting Margin

The Dutch Tweede Kamer passes 96%+ of motions. Pass rate is a structurally useless ceiling-capped metric. Always compute voting_margin as the primary success metric before drawing conclusions about motion effectiveness.

How to compute:

margin = (voor - tegen) / (voor + tegen + afwezig)

This produces a continuous [-1, +1] scale from unanimous rejection to unanimous support. A motion passing 14-1 (margin=+0.87) and one passing 8-7 (margin=+0.07) are both "passed" — the margin exposes the real signal.

Validation: Correlate voting margin against centrist support. If Spearman ρ > 0.7, margin captures a meaningful success gradient that pass rate cannot. In the Dutch data, ρ=0.812.

Implementation: analysis/right_wing/voting_margin.py — reads motions.voting_results (per-party JSON), computes per-motion margin, produces quartile-stratified statistics, and computes pre/post Cohen's d.

Key design decision: Per-party aggregation (1 party = 1 vote) rather than seat-weighted. This measures breadth of cross-spectrum support — exactly the Overton concept — without conflating coalition size effects.

2. Coalition Coding Must Split 2024 at July 1

Treating all 2024 as the Schoof cabinet (PVV/VVD/NSC/BBB) overestimates the coalition effect. The year contains two governments:

• Jan–Jun 2024: Rutte IV (VVD/D66/CDA/CU) — caretaker, limited legislative agenda
• Jul–Dec 2024: Schoof I (PVV/VVD/NSC/BBB) — right-wing coalition

Method: When filtering opposition-only motions, use submission date (not year) against a date-aware coalition dictionary. Motions from Jan–Jun 2024 whose lead submitter is in the Rutte IV coalition must be filtered as government motions, not opposition.

Impact: In the Dutch data, this reclassified 32 PVV/BBB motions from "opposition" to "government," correcting an artificial inflation of the opposition-only shift. Without this fix, the opposition-only Cohen's d overstates the shift by attributing coalition-party motions to opposition.

Implementation: Embedded in all opposition-filter scripts (temporal_trajectory.py, causal_timing.py, predictive_model.py) via COALITION dict. Scripts that only use year-level filtering contain a fragility note.

3. Party Differentiation: Disaggregate Right-Wing Bloc

Right-wing parties are not monolithic. Before attributing any Overton shift to "the right," disaggregate by party to identify the actual driver.

Required dimensions per party:

• Volume (motion count) over time
• Centrist support over time
• Material impact over time
• Pre/post CS delta and volume delta

Method: Parse submitter party from motion title prefixes using regex patterns matching Dutch motion conventions (e.g., "Motie van het lid Wilders ..."). Use mp_metadata table as authority for last-name-to-party mapping. Normalize party names via _PARTY_NORMALIZE (e.g., Groep Markuszower → PVV).

What to look for:

• A party whose centrist support rose without filing more motions: not the driver
• A party whose volume rose without CS increase: also not the driver
• A party with both volume AND CS gains: the primary driver
• A party that entered government (reduced motions, milder content): follower, not leader

Surprising finding from Dutch data: PVV (largest right-wing party) moderated significantly after entering government — fewer, milder motions. JA21 (smaller opposition party) was the primary driver of both volume and support gains. The aggregate "right-wing shift" narrative masks a succession story: PVV moderated, JA21 filled the vacated extreme space.

Implementation: analysis/right_wing/party_differentiation.py — parses submitter parties, builds yearly aggregates, produces 4-panel figure (volume × CS × material impact × pre/post bars).

4. Left-Wing Response Analysis: Verify Asymmetry

An Overton window shift is not symmetric. Rising centrist support for right-wing motions could theoretically be driven by left-wing parties hardening their opposition (making centrist support appear to rise relative to a higher denominator). Always verify the left-wing side by computing centrist support for left-wing motions.

Method:

1. Identify left-wing motions using the same submitter parsing approach
2. Compute centrist support for left motions pre/post
3. Compare Cohen's d for centrist → right shift vs centrist → left shift
4. Compute per-party left-wing softening scores (Δ in CS for each left party)

Interpretation guard: If the centrist shift toward right-wing motions is 10×+ larger than any left-wing hardening, the asymmetry confirms the shift is genuine centrist acceptance, not artifact of denominator change. If both sides shifted, investigate coalition mechanics.

Dutch finding: Centrist shift toward right (d=+1.89) was 18.3× larger than left-wing hardening (d=−0.75). Left-wing opposition support barely moved (21.3%→20.2%). Volt was the only left party that softened (+12.9pp). This extreme asymmetry eliminates denominator artifacts as alternative explanations.

Implementation: analysis/right_wing/left_wing_response.py — loads motions with CANONICAL_LEFT submitter filtering, computes both centrist support and left-party support for right-wing motions, produces asymmetry figures.

5. Mechanism Validation with Second Classifier

LLM-based mechanism classifications are inherently subjective. Always validate inter-rater reliability before drawing conclusions from mechanism distributions.

Method:

1. Classify motion mechanism with primary classifier (inline subagent, reading full text)
2. Run independent second classifier with a different prompt:
   • Use English wording (vs Dutch for original)
   • Present mechanisms in reversed order
   • Include explicit definitions for each mechanism
   • Require strict JSON output with confidence scores
3. Compute Cohen's κ on overlapping classifications
4. For disagreements, prefer the classification with confidence ≥ 4, otherwise retain original

Agreement thresholds:

• κ < 0.2: taxonomy is unusable, scrap and rebuild
• 0.2 ≤ κ < 0.4: taxonomy needs major revision
• 0.4 ≤ κ < 0.6: taxonomy needs targeted refinement (merge ambiguous pairs)
• κ ≥ 0.6: taxonomy is sufficiently reliable

When κ is moderate (0.4–0.6): Analyze the disagreement pairs. In the Dutch data, the primary confusion was institutional_rule_of_law ↔ targeted_restriction — many motions addressing migration enforcement could be classified as either. This revealed a taxonomy ambiguity, not a classification failure: institutional and targeted mechanisms can coexist in the same motion. The fix is not to discard the taxonomy but to note this ambiguity and consider a multi-label classification approach for future iterations.

Implementation: analysis/right_wing/mechanism_validation.py — runs second classifier via parallel LLM calls, computes κ, confusion matrix, resolves disagreements, generates validated distribution.

6. Predictive Model: Feature Importance, Not Just Accuracy

A predictive model's primary value in this domain is not prediction but feature importance — identifying which motion characteristics drive centrist acceptance.

Method:

1. Build features: submitter party (one-hot), category, 2D extremity scores, year, word count, sentiment, coalition membership
2. Target: high centrist support (CS > 0.5) as binary classification
3. Train logistic regression (interpretable coefficients) and random forest (captures non-linearities)
4. Run stratified 5-fold cross-validation with AUC-ROC as primary metric
5. Report top-10 feature importance rankings from both models

What models cannot do: Predict individual motion outcomes in a new parliament. Coalition dynamics, the specific political moment, and the text of the motion interact in ways no tabular model captures. The models validate that the factors we hypothesize as important (party, extremity, category) are indeed predictive — this is methodological triangulation, not forecasting.

Dutch finding: AUC-ROC=0.81 (logistic), 0.84 (RF). Top predictors: submitter party (FVD=−1.33 → hard to get support), SGP (+0.99 → easy), category (asylum/migration hardest), and stijl_extremiteit (higher stylistic extremity predicts lower support). These coefficients validate the analytical framework.

Implementation: analysis/right_wing/predictive_model.py — sklearn logistic regression + random forest, cross-validated, with SHAP-style coefficient interpretation.

7. SVD Spatial Divergence: Center of Gravity Trajectories

The core methodology established Procrustes-aligned SVD drift. The extension decomposes this into group-level center-of-gravity trajectories to measure divergence.

Method:

1. Compute Procrustes-aligned party coordinates per window (from core methodology)
2. For each window, compute the center of gravity for the centrist bloc and right-wing bloc as the component-wise mean
3. Plot both trajectories on the same axes, with year labels
4. Compute Euclidean distance between the two centers of gravity at each window
5. Test for trend: is the inter-bloc distance increasing?

What divergence means: If centrists moved LEFT while right-wing stayed put or moved right, the distance between groups grew. This is acceptance without conversion — the Overton window widened without centrists converting to right-wing positions.

Implementation: analysis/right_wing/svd_trajectory_viz.py — two-panel figure showing individual party arrow trajectories and group center-of-gravity movement.

8. Temporal Trajectory: Quarterly Granularity

Binary pre/post analysis is too coarse to distinguish an electoral shock from a gradual trend. Always decompose into quarters.

Method:

1. Assign each motion to a quarter based on submission date
2. Compute quarterly mean centrist support, with 95% CI
3. Compute quarter-over-quarter Cohen's d to identify the largest single-step change
4. Run a breakpoint detection method (e.g., PELT or binary segmentation) to find structural changes
5. Annotate with known political events (elections, coalition changes, EU-level shifts)

What to look for:

• Immediate jump: the shift was an electoral/political event, not a trend
• Gradual ramp: the shift was a cultural/social trend, harder to attribute to specific events
• Peak and revert: if support peaked and returned to baseline, the shift may be temporary

Dutch finding: The shift was an immediate electoral jump (2023-Q4→2024-Q1: +0.18), suggesting the right-wing electoral victory, not a gradual normalization process. The peak at 2024-Q4 (0.648) followed by reversion to 0.334 by 2026-Q1 raises the possibility of a temporary "coalition honeymoon" rather than permanent shift.

Implementation: analysis/right_wing/temporal_trajectory.py — quarterly binning from submission dates, computes rolling statistics, CIs, and event annotations. causal_timing.py adds political event overlay and breakpoint detection.

9. 2D Extremity Divergence: Separate Stylistic from Material

Single-dimension extremity scores masked diverging trends. Always decompose into stylistic and material dimensions.

Method:

1. Use the two-dimensional rescoring pipeline from the core methodology to score both stijl_extremiteit and materiele_impact
2. Compute yearly means separately for each dimension
3. Test for divergence: are the two dimensions moving in opposite directions?
4. Compute paired Wilcoxon signed-rank test between pre and post periods for each dimension
5. If significant divergence is found, single-score trend analysis is structurally misleading

When to suspect divergence: If the single-score trend is flat (d≈0) but the underlying dimensions are correlated with centrist support in opposite directions, the flat trend masks real change. A decrease in material impact (motions got substantively milder) can cancel out an increase in stylistic extremity (language got sharper), producing a misleading "no change" signal.

Dutch finding: Material impact decreased (−0.146) while stylistic extremity increased (+0.097), with significant Wilcoxon (p=0.002). The single-score "flat trend" was an artifact of these cancelling directions.

Implementation: analysis/right_wing/extremity_2d_temporal.py — joins extremity_scores_2d with right_wing_motions, computes per-dimension yearly means, runs Wilcoxon, generates dual-axis time series figure.

10. Mechanism Classification: Stratified Sampling with Chi-Squared

When testing whether a specific mechanism (e.g., consensus framing) drives centrist support, use stratified sampling and formal hypothesis testing.

Method:

1. Stratify the motion population into 2×2 buckets: period (pre/post) × support level (high/low CS)
2. Sample deterministically (fixed motion IDs) for reproducibility
3. Classify mechanism through manual/subagent review of full title + body text
4. Test H0: mechanism distribution is independent of support level (chi-squared contingency test)
5. Test H0: consensus framing is equally common in high-CS and low-CS post-2024 motions (2×2 Fisher/chi-squared)

Support threshold: CS > 0.5 for "high" support. This threshold is arbitrary but produces clear separation between mechanisms that appeal broadly (consensus framing) vs those that polarize (targeted restriction, system dismantling).

When to reject the taxonomy: If κ < 0.4 from independent validation, the mechanism labels themselves are unreliable and chi-squared results are misleading. In this case, report κ rather than mechanism findings.

Dutch finding: Consensus framing accounts for 24% of high-CS post-2024 motions vs 8% of low-CS (significant). This validates the mechanism hypothesis even with moderate κ=0.41 — the signal survives the noise.

Implementation: analysis/right_wing/mechanism_classification.py — deterministic sample of 200 motions, inline classifications, chi-squared tests, generated report.

Why This Matters

Without these extensions, Overton window analysis produces results that are correct in direction but misleading in detail:

• Binary pass rate creates false nulls (no detectable change)
• Undifferentiated "right-wing" analysis misattributes effects to the wrong party
• Missing left-wing verification cannot rule out denominator artifacts
• Single-dimension extremity masks opposite trajectories in stylistic vs material content
• Missing mechanism validation overstates confidence in classification-based findings
• Annual aggregation cannot distinguish shock vs trend

When to Apply

• When pass rate exceeds 90% and provides no signal (always in multi-party parliaments with high passage rates)
• When right-wing parties have heterogeneous trajectories — some entering government, some remaining in opposition
• When the analysis must rule out left-wing denominator effects as alternative explanations
• When LLM-scored extremity metrics are used as independent variables
• When mechanism classification forms part of the explanatory framework
• When temporal resolution matters: quarterly decomposition is necessary to distinguish electoral shock from gradual shift
• When 2D extremity data is available and single-score trends appear flat

Related

• overton-window-shift-methodology-2026-05-24.md — core 7-step methodology (prerequisite)
• domain-decomposition-overton-analysis.md — per-domain decomposition (apply before interpretation)
• svd-labels-voting-patterns-not-semantics.md — foundational: SVD captures voting patterns
• .opencode/skills/score-extremity/SKILL.md — two-dimensional extremity scoring pipeline