Methodology

In the first iteration of this analysis, we introduced the notion of a “root map”—the most population compact map for a state. Unfortunately, we found that the root map was not a good reference point for comparing human-drawn maps, because it can’t be computed definitively (only approximated) and small changes in the approximation yielded wildly different ratings. To use the metaphor of a landscape, rather than getting deeper and deeper into the valley containing the global minimum, successive approximations were flipping between different valleys with local minima and very different characteristics.

In this iteration, we’ve take a different approach. In summary:

• We start with the data, shapes, and contiguity graphs in the RDA Base repository. The census (total population) and demographic data (voting age populations by race and ethnicity) are from the 2020 Census (unadjusted). The election data are the 2016-2020 election composite from Dave’s Redistricting’s (DRA) public VTD data. DRA got most of the election data from data partners, especially Voting and Election Science Team (VEST).
• We generate an unbiased ensemble of 10,000 congressional plans for a state, using the ReCom algorithm implemented in GerryChain. These plans all use whole precincts, i.e., don’t split precincts with subsets of blocks in different districts.
• Then we score each plan in the ensemble on the five ratings dimensions in DRA, using a high-volume offline scoring service which wraps the Python clone of DRA analytics. Note: We removed the clipping of minority VAP percentages below 37% that DRA’s minority rating does, because this was causing a discontinuity (discernible gaps) in the ratings.
• Next we find the Pareto frontier of the plans in the ensemble that satisfy the “realistic” constraints of DRA’s the Notable Maps.
• Then we export notable maps for the state from DRA and pull their ratings. If necessary, we make close-cousin clones of these maps that don’t split any precincts for use below.
• These human-drawn plans each optimize for one of the five rating dimension, and we use them as seeds to GerryChain’s SingleMetricOptimizer and Gingelator exploratory (biased) chain algorithms.
• We use each as the starting point for seeking better-rated plans on each of the five dimensions, one at a time. We use 2,000 iterations for each optimization run which generates another 10,000 plans for each seed. However, we only retain interim plans that improve the metric for which we are optimizing.
• To understand this visually, think of a pair ratings dimensions, say proprtionality and compactness. Think of a scatter plot of proportionality (y-axis) vs. compactness (x-axis). The most proportional map will plot somewhere in the upper left of the plot; the most compact map will plot somewhere in the lower right. By using those starting points and optimizing away from them along the other ratings dimensions, we, in essence, search for conforming plans upper-left to lower-right and vice versa.
• We optimize for each of these dimensions, using proxy metrics as opposed to the actual DRA ratings: for proportionality, compactness, and county-district splitting, we use the efficiency gap, Polsby-Popper, and the number of county splits, respectively, all of which are built into Gerrychain; for minority opportunity, we use Gerrychain’s Gingelator feature; and for competitiveness, we use the raw metric that DRA uses before normalizing it to [0-100].
• We collect those plans into a new optimized ensemble and score them.
• We combine the unbiased ensemble and the optimized ensemble into a single ensemble and find the new Pareto frontier, again filtering for realistic plans.
• Finally, we generate all the artifacts that you see on the state pages, e.g., for NC.

Besides the GitHub repositories mentioned above, the main code and data for this workflow is in three repositories: rdabase, rdaensemble, and tradeoffs, the last being where this site is hosted. Two more repositories—rdadccvt and rdaroot—package up our work on root maps, which we use as seeds to the unbiased ensembles here.