Another Gerrymandering Game

Here's a game to practice your gerrymandering by packing and cracking. It starts pretty simple on a 3x3 board making districts of 3 blocks and builds up to a 7x7 board building 7 block districts. I am frequently a completionist on little games like this, but when you get to level 50 it will just keep going forever, so when you get there, congratulations, you won the game.


Partisan Outcome of Compact Districts

I have often been asked what the likely partisan breakdown of my maps would be. I never had the data to do that analysis, but FiveThirtyEight got the data and analyzed my maps and 7 other plans for various notions of gerrymandered and fair.
In short, my compact maps are very slightly less Republican and a bit more competitive. I've always believed that on average in the last two districtings Republicans have done more of the gerrymandering and stolen more US House seats than the Democrats have. The difference was smaller than I expected though. The difference is probably smaller than this model can usefully tell us about. All models are wrong, some models are useful, and this model is certainly not destiny but can maybe tell us something useful about the biases in the system.
FiveThirtyEight estimates that the current map has 195 safe Republican seats, 168 safe Democratic seats, and 72 competitive seats; and that this will on average elect 234.4 Republicans and 200.6 Democrats. Actually in 2016 we got 241 Republicans and 194 Democrats. 7 seats out of 435, 1.6% off, not bad.
They estimate that under my compact map that goes to 180 safe Republican seats (-15), 151 safe Democratic seats (-17), and 104 competitive seats (+32); with an expected outcome of 232.2 R and 202.8 D (D +2.2).
That 2.2 seat change looks pretty small, but I want to be optimistic about the 32 additional competitive districts. The US may be self-sorting, but while gerrymandering deliberately creates uncompetitive districts on both sides, simply not gerrymandering creates additional competitive districts. I'm not in favor of distorting districts specifically to create competitive districts (which 538 explored and created a whopping 242 competitive districts) but not deliberately creating uncompetitive districts is something I'm solidly behind. Uncompetitive districts make democracy depressing when you know you don't really have any choice, and I want this country to have more democracy.


Pennsylvania Fair District Rules

In a decision handed down on January 22, 2018 by the Pennsylvania Supreme Court, they decreed that gerrymandered districts shall be thrown out and replaced such that (emphasis mine):
any congressional districting plan shall consist of: congressional districts composed of compact and contiguous territory; as nearly equal in population as practicable; and which do not divide any county, city, incorporated town, borough, township, or ward, except where necessary to ensure equality of population. 
I stared at this paragraph for a boggled minute because the English clauses aren't clearly laid out in a prioritized order. Rationalizing it combined with what I know of redistricting law, I think the priority has to work out to be:

  1. contiguous
  2. equal population
  3. whole towns
  4. compact
If that is a complete statement of what the Pennsylvania courts want out of a district map, it's relatively simple. I clearly need to get hacking on the municipality-preserving version of my compact district solver.


Brown CS k-means Redistricting

K-Means redistricting at Brown U

I tried this 10 years ago and it's fast and simple but it isn't good enough.

The solver in the article seems to completely ignore existing boundaries including the boundaries to which the Census has actually counted data called 'blocks' (sometimes a city block, sometimes an empty square mile of Montana). If they're assuming that population is fungible and uniformly distributed within a Census block I think that's invalid. My k-means solver drew those nice straight lines but then assigned whole blocks based on whether the center of the block was one side of the line or the other. All the districts thus had ragged edges.

Because the k-means algorithm has very few data points to fiddle (district center, district weight) it can't find complex solutions. When I ran a k-means solver it couldn't find districts with close enough to equal population in a few places.

I eventually went with a solver that considered one at a time each block on a border of two districts to see if a block would be better moved to the other district. That kind of detail and flexibility made maps that weren't as ideally simple, but still had good compactness, had much better equal-population constraint satisfaction, and should be much more workable for observing the same Census blocks that existing redistricting is done on.


Sotomayor on Gerrymandering's Absurdity

"Could you tell me what the value is to democracy from political gerrymandering?
How -- how does that help our system of government?"
"... it's okay to stack the decks so that for 10 years or an indefinite period of time one party, even though it gets a minority of votes, can't get a minor -- gets a minority of votes, can get the majority of seats?"
 -- Justice Sonia Sotomayor

from oral arguments before the Supreme Court of the United States in the case of Gill v Whitford (on the matter of gerrymandering in Wisconsin)


How the SCOTUS Wisconsin Gerrymander case could go

A. Partisan gerrymander bad! Must gerrymander districts to ensure shoddy proportional representation.
B. Partisan gerrymander bad! Map must provably have no bias.
C. Meh. Whatever. Let whatever party do whatever they want.
D. BS! Puny anti-gerrymandering arguments are so bad we don’t want any courts to listen to them in 2021 either.

I feel like the plaintiffs want A. I feel like this is pretty unlikely. It would be a truly revolutionary decision that would effectively create new rules about how every state needs to redistrict. The Court is loathe to do this. Congress could pass a new 'Voting Rights Act' type law which could do this, if Congress could pass anything useful.

Option B would be the court pulling a rabbit out of a hat. I'm not sure anyone is asking for this except me.

I would mostly bet on C. No change, no big decision either way.

I am afraid of D. The court could make a ruling such that they not only throw out the Wisconsin complaint but throw out the basis and statistical means of the Wisconsin complaint thus removing the basis for future court challenges. I know a lot of people who are betting on playing a good court challenge game in 2021 to battle the expected gerrymanders which come after the next Census. There would be a lot of searching for a new strategy between now and then if SCOTUS actively disqualifies the tests being tried in the Wisconsin case.


Things are in the news now, we can read snippets of the arguments in Court, but there won't be real news until the justices make their decision in the coming months.


Compute budget

I heard of a computational anti-gerrymandering project with a 6,000,000 cpu-hour budget on a university cluster supercomputer. I worked out that my main run in 2011 was a bit under 20,000 cpu-hours. My project at current prices with AWS spot instances or GCE interruptible instances that works out to $130-$150. That's a bit cheaper than what I paid for my home computer, and I could parallelize and do all the compute in a week instead of stringing it out over six months.