Untangling the Dueling Expert Witnesses: Comparing Ensemble Methods in Pennsylvania's Redistricting Plans

TL;DR

This paper compares ensemble methods for analyzing Pennsylvania's legislative districts, highlighting discrepancies in expert witness testimonies and advocating for transparent, election-based metrics over proprietary data.

Contribution

It introduces a comparative analysis of ensemble methods like Redist and Gerrychain with constraints, and critiques the use of vote summation metrics in gerrymandering assessments.

Findings

Ensemble methods can produce differing results in district analysis.

Vote summation metrics are unreliable for gerrymandering detection.

Transparency is improved when using publicly available election data.

Abstract

Ensembles of random legislative districts are a valuable tool for assessing whether a proposed district plan is an outlier or gerrymander. Expert witnesses have presented these in litigation using various methods, and unsurprisingly, they often disagree. Recent open source methods now permit independent validation of expert witness testimony. Here, we compare ensembles for the Pennsylvania House and Congressional districts calculated using "Redist" and "Gerrychain" further incorporating constraints restricting county and municipal boundary splitting, as required by Pennsylvania for legal plans. We compare results to expert witness testimony submitted by Republican and Democratic parties. We confirm some of the testimony but could not reproduce all of it, struggling with metrics based on a heuristic "sum of votes index" rathern than a straightforward average of metrics across multiple elections. We recommend against relying on analytics based on summing votes from multiple elections to create vote incides and derivative metrics as these are inherently poorly behaved. To promote transparency, we recommend that where possible, expert witness testimony be based solely on publicly available election data as opposed to proprietary data closely held by political parties.