Don’t Trust a Single Gerrymandering Metric

Abstract In recent years, in an effort to promote fairness in the election process, a wide variety of techniques and metrics have been proposed to determine whether a map is a partisan gerrymander. The most accessible measures, requiring easily obtained data, are metrics such as the Mean-Median Difference, Efficiency Gap, Declination, and GEO metric. But for most of these metrics, researchers have struggled to describe, given no additional information, how a value of that metric on a single map indicates the presence or absence of gerrymandering. Our main result is that each of these metrics is gameable when used as a single, isolated quantity to detect gerrymandering (or the lack thereof). That is, for each of the four metrics, we can find district plans for a given state with an extremely large number of Democratic-won (or Republican-won) districts while the metric value of that plan falls within a reasonable, predetermined bound. We do this by using a hill-climbing method to generate district plans that are constrained by the bounds on the metric but also maximize or nearly maximize the number of districts won by a party. In addition, extreme values of the Mean-Median Difference do not necessarily correspond to maps with an extreme number of districts won. Thus, the Mean-Median Difference metric is particularly misleading, as it cannot distinguish more extreme maps from less extreme maps. The other metrics are more nuanced, but when assessed on an ensemble , none perform substantially differently from simply measuring number of districts won by a fixed party. One clear consequence of these results is that they demonstrate the folly of specifying a priori bounds on a metric that a redistricting commission must meet in order to avoid gerrymandering.