Abstract
Traditional gerrymandering detection tools often rely on two-party assumptions that fail in multiparty systems. This paper proposes a novel nonparametric statistical method using kernel regression to "learn" the relationship between votes and seats from empirical data. By conditioning expectations on the effective number of competitors, the method identifies electoral bias as a statistical anomaly. Validated on both simulated and real-world datasets, including U.S. and Polish elections, the approach demonstrates high diagnostic power and flexibility across diverse electoral landscapes.
Introduction
The detection of gerrymandering—the intentional manipulation of electoral district boundaries to favor a specific party—is a cornerstone of maintaining democratic integrity. However, established metrics like the "efficiency gap" or "partisan symmetry" are primarily tailored for the U.S. two-party system. These tools frequently break down in multiparty contexts or when parties do not contest every district.
The fundamental challenge lies in the non-linear translation of votes into seats in "First-Past-The-Post" (plurality) systems. This translation is influenced by the concentration of supporters and the number of competitors. Distinguishing between "natural" bias (e.g., urban clustering) and intentional "packing and cracking" requires a model that can adapt to varying electoral conditions.
Methodology
The authors frame gerrymandering as a statistical anomaly within the vote-to-seat translation process. Their nonparametric framework rests on three technical pillars:
- Effective Number of Competitors ($\phi$): To account for multiparty complexity, the researchers employ a metric derived from the Herfindahl-Hirschman-Simpson index. This quantifies the competitive landscape of a district, regardless of the absolute number of candidates.
- Effective Seat Thresholds: Using Support Vector Machines (SVM) and B-splines, the model identifies the "decision boundary"—the specific vote share required for a candidate to have a 50% probability of winning, given the local competition pattern.
- Locally-Constant Kernel Regression: Instead of prescribing a theoretical "fair" curve (like the "Cube Law"), the method uses Nadaraya-Watson kernel regression to learn the expected seat share from a large training set of historical elections.
By calculating an Electoral Bias p-Value ($\pi_i$), the system flags parties whose seat shares significantly deviate from the learned norm as potential victims or beneficiaries of gerrymandering.
Data and Empirical Analysis
The model was trained and tested on a diverse array of data:
- Simulated Plans: 192 "fair" plans (via Monte Carlo algorithms) and 128 "unfair" gerrymandered plans (via integer linear programming) modeled on the geographies of France, Germany, and Poland.
- Empirical Data: Polish municipal elections (2014, 2018), U.S. House elections (1900–2022), and national legislative data from 15 countries via the Constituency-Level Elections Archive.
Results and Discussion
The method demonstrated significant diagnostic power. In experimental settings with known ground truth, it achieved a precision of 0.896 and a recall of 0.912.
When applied to 2022 U.S. Congressional data, the kernel method's results showed a strong correlation (0.623) with the aggregate of eight other major gerrymandering indices. This indicates that the nonparametric approach successfully captures the essence of gerrymandering even in the two-party context it wasn't specifically designed for. The primary insight is that fairness can be learned from data rather than defined a priori, allowing the model to account for the unique "skewness" of different political geographies.
Conclusion
While the proposed method is highly effective at identifying electoral bias, the authors note the difficulty in distinguishing between intentional gerrymandering and natural geographic bias (e.g., urban voter concentration). Future research will focus on refining benchmarks through purely random election simulations and investigating discrepancies between this statistical approach and traditional geometric district-shape indices.
Summary updated