Auditing for Gerrymandering by Identifying Disenfranchised Individuals
Gerrymandering is the practice of drawing congressional districts to advantage or disadvantage particular electoral outcomes or population groups. We study the problem of computationally auditing a districting for evidence of gerrymandering. Our approach is novel in its emphasis on identifying individual voters disenfranchised by packing and cracking in local fine-grained geographic regions. We define a local score based on comparison with a representative sample of alternative districtings and use simulated annealing to algorithmically generate a witness districting to show that the score can be substantially reduced by simple local alterations. Unlike commonly studied metrics for gerrymandering such as proportionality and compactness, our framework is inspired by the legal context for voting rights in the United States. We demonstrate the use of our framework to analyze the congressional districting of the state of North Carolina in 2016. We identify a substantial number of geographically localized disenfranchised individuals, mostly Democrats in the central and north-eastern parts of the state. Our simulated annealing algorithm is able to generate a witness districting with a roughly 50% reduction in the number of disenfranchised individuals, suggesting that the 2016 districting was not predetermined by North Carolina's spatial structure.
